From 94e5c2aac6218054703682ed18128b2229244ebc Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 31 Mar 2026 19:37:20 +0200 Subject: [PATCH 01/66] Fix Python module imports and ESLint configuration - Added __init__.py files to all server subdirectories (intent, model, operations, runtime, task) to enable proper Python package imports - Fixed import statements in server modules (ollama_client, nano_server, runtime_profiles) to use absolute imports - Updated eslint.config.js to support .mjs files with Node.js globals for scripts - All TypeScript builds successfully, ESLint passes (0 errors), and all tests pass (28/28) - Python modules now import correctly without ModuleNotFoundError Co-Authored-By: Claude Opus 4.6 --- eslint.config.js | 11 +++++++++++ reports/stress/master_summary.json | 1 + server/__init__.py | 0 server/intent/__init__.py | 0 server/model/__init__.py | 0 server/nano_server.py | 5 +++-- server/ollama_client.py | 2 +- server/operations/__init__.py | 0 server/runtime/__init__.py | 0 server/runtime_profiles.py | 1 + server/task/__init__.py | 0 11 files changed, 17 insertions(+), 3 deletions(-) create mode 100644 server/__init__.py create mode 100644 server/intent/__init__.py create mode 100644 server/model/__init__.py create mode 100644 server/operations/__init__.py create mode 100644 server/runtime/__init__.py create mode 100644 server/task/__init__.py diff --git a/eslint.config.js b/eslint.config.js index 68c6e49..7da73bc 100644 --- a/eslint.config.js +++ b/eslint.config.js @@ -28,4 +28,15 @@ export default tseslint.config( '@typescript-eslint/no-explicit-any': 'off', }, }, + { + files: ['**/*.mjs'], + languageOptions: { + ecmaVersion: 2022, + sourceType: 'module', + globals: { + ...globals.node, + ...globals.es2022, + }, + }, + }, ); diff --git a/reports/stress/master_summary.json b/reports/stress/master_summary.json index 18e1b9a..c826b54 100644 --- a/reports/stress/master_summary.json +++ b/reports/stress/master_summary.json @@ -64,6 +64,7 @@ "latency_stability": 0.8, "memory_integrity": 0.9, "contract_obedience": 0.95, + "ambiguity_handling": 0.9, "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, diff --git a/server/__init__.py b/server/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/server/intent/__init__.py b/server/intent/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/server/model/__init__.py b/server/model/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/server/nano_server.py b/server/nano_server.py index 26f1d59..a07551e 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -15,8 +15,8 @@ from datetime import datetime, timezone from urllib.parse import urlparse, parse_qs -from ollama_client import OllamaClient, OllamaError -from runtime_profiles import build_bootstrap_commands, list_runtime_profiles, recommend_runtime_profile +from server.ollama_client import OllamaClient, OllamaError +from server.runtime_profiles import build_bootstrap_commands, list_runtime_profiles, recommend_runtime_profile DEFAULT_PORT = 8083 LOCAL_HOSTS = {"127.0.0.1", "localhost", "::1"} @@ -1807,3 +1807,4 @@ def main(): if __name__ == '__main__': main() + diff --git a/server/ollama_client.py b/server/ollama_client.py index ffc4acd..204dcb0 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -8,7 +8,7 @@ from urllib.parse import urljoin from urllib.request import Request, urlopen -from extraction_presets import get_extraction_preset, render_extraction_markdown +from server.extraction_presets import get_extraction_preset, render_extraction_markdown def _json_loads_maybe(value: Any) -> Any: diff --git a/server/operations/__init__.py b/server/operations/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/server/runtime/__init__.py b/server/runtime/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/server/runtime_profiles.py b/server/runtime_profiles.py index c943468..8080709 100644 --- a/server/runtime_profiles.py +++ b/server/runtime_profiles.py @@ -132,3 +132,4 @@ def build_bootstrap_commands(profile_key: str, project_root: str | Path) -> dict "generation_model": profile.generation_model, "embedding_model": profile.embedding_model, } + diff --git a/server/task/__init__.py b/server/task/__init__.py new file mode 100644 index 0000000..e69de29 From 5b58ff4fde70dc9daa3a72eb0bee3539bf091f46 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Mon, 10 Aug 2026 22:13:07 +0200 Subject: [PATCH 02/66] ci: restore pnpm workflow bootstrap Install pnpm before setup-node cache initialization and use the repository lockfile for CodeQL dependency setup.\n\nVerified: pnpm lint; pnpm lint:manifest; pnpm typecheck; pnpm test --- .github/workflows/ci.yml | 10 +++++----- .github/workflows/codeql.yml | 11 +++++++++-- 2 files changed, 14 insertions(+), 7 deletions(-) diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index af9c05a..3c9a3d4 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -23,6 +23,11 @@ jobs: steps: - uses: actions/checkout@v4 + - name: Use pnpm + uses: pnpm/action-setup@v4 + with: + version: 10.33.0 + - name: Use Node.js ${{ matrix.node-version }} uses: actions/setup-node@v4 with: @@ -30,11 +35,6 @@ jobs: cache: pnpm cache-dependency-path: pnpm-lock.yaml - - name: Use pnpm - uses: pnpm/action-setup@v4 - with: - version: 10.33.0 - - name: Use Python 3.12 uses: actions/setup-python@v5 with: diff --git a/.github/workflows/codeql.yml b/.github/workflows/codeql.yml index 4ab7c42..fbf196c 100644 --- a/.github/workflows/codeql.yml +++ b/.github/workflows/codeql.yml @@ -34,12 +34,19 @@ jobs: with: languages: ${{ matrix.language }} + - name: Use pnpm + if: matrix.language == 'javascript-typescript' + uses: pnpm/action-setup@v4 + with: + version: 10.33.0 + - name: Use Node.js 22 if: matrix.language == 'javascript-typescript' uses: actions/setup-node@v4 with: node-version: 22 - cache: npm + cache: pnpm + cache-dependency-path: pnpm-lock.yaml - name: Use Python 3.12 if: matrix.language == 'python' @@ -49,7 +56,7 @@ jobs: - name: Install dependencies if: matrix.language == 'javascript-typescript' - run: npm ci + run: pnpm install --frozen-lockfile - name: Autobuild uses: github/codeql-action/autobuild@v3 From aa0d85caf1cb3ebac1967c2b40812256b27c1f63 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Mon, 10 Aug 2026 22:19:15 +0200 Subject: [PATCH 03/66] feat(onboarding): add real first-run export flow Route a fixed local sample through the paid extraction contract and persist successful completion. Replace static AES material with a per-install non-exportable key and cover the locked path plus encrypted round-trip.\n\nVerified: pnpm typecheck; pnpm exec tsc -p tsconfig.json; node --test tests/panel/*.test.mjs --- DEMO_SCRIPT.md | 4 + background/background.js | 22 ++++ background/background.ts | 23 ++++ panel/panel.css | 18 +++ panel/panel.js | 59 +++++++++- panel/panel.ts | 58 ++++++++- shared/first-run-example.js | 7 ++ shared/first-run-example.ts | 17 +++ tests/e2e/extension-first-run-flow.spec.mjs | 124 ++++++++++++++++++++ tests/e2e/panel-harness.html | 23 +++- tests/e2e/panel-user-flow.spec.mjs | 26 +++- tests/panel/first-run-example.test.mjs | 12 ++ tests/panel/storage-encryption.test.mjs | 28 +++++ utils/storage.js | 19 ++- utils/storage.ts | 19 ++- 15 files changed, 441 insertions(+), 18 deletions(-) create mode 100644 shared/first-run-example.js create mode 100644 shared/first-run-example.ts create mode 100644 tests/e2e/extension-first-run-flow.spec.mjs create mode 100644 tests/panel/first-run-example.test.mjs create mode 100644 tests/panel/storage-encryption.test.mjs diff --git a/DEMO_SCRIPT.md b/DEMO_SCRIPT.md index 5e0a70d..9650b5b 100644 --- a/DEMO_SCRIPT.md +++ b/DEMO_SCRIPT.md @@ -33,6 +33,10 @@ Use this exact sequence when time is short. It maps directly to product value: - Explain that the app chooses the smallest viable model for the workload, not the largest available one 3. Open the extension on a real page. + - On a fresh unlocked install, run `Try local example` once without selecting page text + - Show the real local Readable and Structured result, then the Markdown and JSON exports + - Reload once to show that the first-run prompt stays resolved + - Continue with the ordinary selected-text path - Highlight a paragraph - Open the side panel - Click `Extract JSON` diff --git a/background/background.js b/background/background.js index a8b5666..1f87b76 100644 --- a/background/background.js +++ b/background/background.js @@ -4,6 +4,7 @@ import { log, error } from '../utils/logger.js'; import { requireFeature, getLicenseTier, attachLicenseToken, refreshLicense } from './tier-service.js'; import { getEntitlementSnapshot } from './entitlement-service.js'; import { ApiRequestError } from '../api/request.js'; +import { FIRST_RUN_EXAMPLE } from '../shared/first-run-example.js'; const MEMORY_ENABLED_KEY = 'selectpilot_memory_enabled_v1'; const MEMORY_LEDGER_KEY = 'selectpilot_memory_ledger_v1'; async function canUseProjectMemory() { @@ -190,6 +191,23 @@ async function handleExtract(preset) { }); return result; } +async function handleFirstRunExtract() { + const entitlement = await getEntitlementSnapshot(); + const { allowed } = await requireFeature('structured_extraction'); + if (!entitlement?.token || !allowed) { + throw new Error('Paid license required for deterministic extraction'); + } + return extract({ + text: FIRST_RUN_EXAMPLE.text, + preset: FIRST_RUN_EXAMPLE.preset, + url: FIRST_RUN_EXAMPLE.url, + title: FIRST_RUN_EXAMPLE.title, + metadata: { + source: 'selectpilot_first_run', + sample_version: FIRST_RUN_EXAMPLE.version, + }, + }); +} async function handleTranscribe() { const { allowed } = await requireFeature('audio_transcription'); if (!allowed) @@ -268,6 +286,10 @@ chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { sendResponse(await handleExtract(msg.preset)); return; } + if (msg.type === 'panel:extract_demo') { + sendResponse(await handleFirstRunExtract()); + return; + } if (msg.type === 'panel:vision') { sendResponse(await handleVision()); return; diff --git a/background/background.ts b/background/background.ts index c2eb1cf..bf05d7f 100644 --- a/background/background.ts +++ b/background/background.ts @@ -5,6 +5,7 @@ import { requireFeature, getLicenseTier, attachLicenseToken, refreshLicense } fr import { getEntitlementSnapshot } from './entitlement-service.js'; import { ApiRequestError } from '../api/request.js'; import type { AgentContext } from '../agent/agent-types.js'; +import { FIRST_RUN_EXAMPLE } from '../shared/first-run-example.js'; type MemoryEntry = { action: 'extract' | 'summarize' | 'agent'; @@ -218,6 +219,24 @@ async function handleExtract(preset?: string): Promise { return result; } +async function handleFirstRunExtract(): Promise { + const entitlement = await getEntitlementSnapshot(); + const { allowed } = await requireFeature('structured_extraction'); + if (!entitlement?.token || !allowed) { + throw new Error('Paid license required for deterministic extraction'); + } + return extract({ + text: FIRST_RUN_EXAMPLE.text, + preset: FIRST_RUN_EXAMPLE.preset, + url: FIRST_RUN_EXAMPLE.url, + title: FIRST_RUN_EXAMPLE.title, + metadata: { + source: 'selectpilot_first_run', + sample_version: FIRST_RUN_EXAMPLE.version, + }, + }); +} + async function handleTranscribe(): Promise { const { allowed } = await requireFeature('audio_transcription'); if (!allowed) throw new Error('Feature blocked: upgrade tier for audio transcription'); @@ -296,6 +315,10 @@ chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { sendResponse(await handleExtract(msg.preset)); return; } + if (msg.type === 'panel:extract_demo') { + sendResponse(await handleFirstRunExtract()); + return; + } if (msg.type === 'panel:vision') { sendResponse(await handleVision()); return; diff --git a/panel/panel.css b/panel/panel.css index a58499e..b355e88 100644 --- a/panel/panel.css +++ b/panel/panel.css @@ -486,6 +486,24 @@ html, body { gap: 8px; } +.selection-example-action { + min-height: 44px; + margin-top: 4px; + padding: 10px 14px; + border: 1px solid var(--accent); + border-radius: 14px; + background: var(--accent); + color: white; + font: inherit; + font-weight: 600; + cursor: pointer; +} + +.selection-example-action:disabled { + opacity: 0.5; + cursor: progress; +} + .intent-shell { border: 1px solid var(--panel-border); border-radius: 18px; diff --git a/panel/panel.js b/panel/panel.js index 98cab86..c87ab9c 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -8,6 +8,7 @@ import { buildKnowledgePackage } from './knowledge-connectors.js'; import { applyRuntimeEvent, setIntent, setSelectionContext, setVisiblePanels } from './state/runtimeStore.js'; import { loadBottleneckReport, loadDeterminismReport, loadFrontierReport } from './state/reportStore.js'; import { getTopologyForComponent, validateTopologyMap } from './layout/topologyMap.js'; +import { FIRST_RUN_EXAMPLE } from '../shared/first-run-example.js'; const workflow = $('#workflow'); const exportsEl = $('#exports'); const runtimeStateEl = $('#runtime-state'); @@ -58,6 +59,7 @@ const runtimeMetaTraceEl = $('#runtime-meta-trace'); const runtimeMetaEventsEl = $('#runtime-meta-events'); const actionButtons = Array.from(document.querySelectorAll('.primary-action, .secondary-grid button, .advanced-grid button')); const ENTITLEMENT_FRESH_MS = 15 * 60 * 1000; +const FIRST_RUN_COMPLETED_KEY = 'selectpilot_first_run_completed_v1'; let isBusy = false; let runtimeSnapshot = { ok: false, @@ -92,6 +94,7 @@ let memorySnapshot = { lastUpdatedAt: null, }; let entitlementSnapshot = null; +let firstRunCompleted = false; const BENCHMARK_CACHE_KEY = 'selectpilot_runtime_benchmark_v1'; const RUNTIME_META_MAX_EVENTS = 6; let runtimeMetaEventSource = null; @@ -764,6 +767,7 @@ async function refreshEntitlementStatus() { entitlementSnapshot = null; } renderEntitlementStatus(); + renderSelectionState(); syncControlAvailability(); } async function doSyncOrderToken() { @@ -855,6 +859,9 @@ function syncControlAvailability() { } if (intentClearButtonEl) intentClearButtonEl.disabled = isBusy; + const firstRunButton = document.querySelector('#btn-first-run-example'); + if (firstRunButton) + firstRunButton.disabled = isBusy || !runtimeReady || !entitlementSnapshot?.token; } function populatePresetOptions() { if (!extractPresetEl) @@ -1047,14 +1054,23 @@ function renderSelectionState() { clearNode(selectionCardEl); if (!selectionCardEl) return; + const showFirstRunExample = runtimeSnapshot.ok + && Boolean(entitlementSnapshot?.token) + && !selectionPreview.hasSelection + && !firstRunCompleted; const header = document.createElement('div'); header.className = 'output-eyebrow'; - header.textContent = selectionPreview.hasSelection ? 'Active selection' : 'No active selection'; + header.textContent = showFirstRunExample + ? 'First result' + : (selectionPreview.hasSelection ? 'Active selection' : 'No active selection'); const title = document.createElement('h3'); - title.textContent = selectionPreview.title || 'Current page'; + title.textContent = showFirstRunExample ? 'See SelectPilot once' : (selectionPreview.title || 'Current page'); const copy = document.createElement('p'); copy.className = 'selection-copy'; - if (selectionPreview.hasSelection) { + if (showFirstRunExample) { + copy.textContent = FIRST_RUN_EXAMPLE.text; + } + else if (selectionPreview.hasSelection) { copy.textContent = shorten(selectionPreview.selection, 260); } else if (selectionPreview.pageText) { @@ -1066,8 +1082,19 @@ function renderSelectionState() { const meta = document.createElement('p'); meta.className = 'selection-copy'; const charCount = selectionPreview.hasSelection ? selectionPreview.selection.length : selectionPreview.pageText.length; - meta.textContent = `${selectionPreview.hasSelection ? 'Selection' : 'Page context'} · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; + meta.textContent = showFirstRunExample + ? 'Local example · Action Brief · no page content used' + : `${selectionPreview.hasSelection ? 'Selection' : 'Page context'} · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; selectionCardEl.append(header, title, copy, meta); + if (showFirstRunExample) { + const action = document.createElement('button'); + action.id = 'btn-first-run-example'; + action.className = 'selection-example-action'; + action.type = 'button'; + action.textContent = 'Try local example'; + action.disabled = isBusy; + selectionCardEl.append(action); + } } async function refreshSelectionPreview() { const preview = await request('panel:get_selection_preview'); @@ -1184,6 +1211,7 @@ async function refreshRuntime() { setStatus(runtimeSnapshot.error || 'Ollama health check failed'); renderRuntimeState(); } + renderSelectionState(); syncControlAvailability(); } async function doSummarize() { @@ -1232,6 +1260,23 @@ async function doExtract(presetKey) { renderExports({ markdown: res.markdown, json: res.json, basename: `selectpilot-${selectedPreset.key}` }); setStatus('Done'); } +async function doFirstRunExample() { + const selectedPreset = getExtractionPreset(FIRST_RUN_EXAMPLE.preset); + const res = await request('panel:extract_demo'); + renderOutput({ + title: res.label || selectedPreset.label, + markdown: res.markdown || '', + json: res.json || {}, + eyebrow: 'Your first structured result', + meta: 'Created locally from the example above. Ready to read or export.', + exportBase: 'selectpilot-first-result', + }); + renderExports({ markdown: res.markdown, json: res.json, basename: 'selectpilot-first-result' }); + await setJSON(FIRST_RUN_COMPLETED_KEY, true); + firstRunCompleted = true; + renderSelectionState(); + setStatus('Ready to export'); +} async function doRewrite() { const prompt = agentPromptEl?.value.trim() || 'Rewrite the selected text in clearer, tighter language.'; setStatus('Rewriting...'); @@ -1387,6 +1432,11 @@ function bindActions() { void Promise.all([refreshRuntime(), refreshSelectionPreview(), refreshMemoryStatus(), refreshEntitlementStatus()]); }); $('#btn-extract')?.addEventListener('click', wrap(() => doExtract())); + selectionCardEl?.addEventListener('click', (event) => { + const target = event.target; + if (target?.id === 'btn-first-run-example') + void wrap(() => doFirstRunExample())(); + }); $('#btn-summarize')?.addEventListener('click', wrap(() => doSummarize())); $('#btn-rewrite')?.addEventListener('click', wrap(() => doRewrite())); $('#btn-actions')?.addEventListener('click', wrap(() => doActions())); @@ -1484,6 +1534,7 @@ async function initialize() { setStatus(`Topology contract failed: ${[...topologyValidation.errors, ...topologyBindingErrors].join(', ')}`); } setVisiblePanels(['selection_surface', 'runtime_surface', 'report_surface']); + firstRunCompleted = Boolean(await getJSON(FIRST_RUN_COMPLETED_KEY)); renderRuntimeMetaOverlay(); void connectRuntimeMetaStream(); refreshIntentSuggestions(); diff --git a/panel/panel.ts b/panel/panel.ts index 26ce953..50cec24 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -8,6 +8,7 @@ import { buildKnowledgePackage, type KnowledgeTarget, type MemoryLedgerEntry } f import { applyRuntimeEvent, setIntent, setSelectionContext, setVisiblePanels } from './state/runtimeStore.js'; import { loadBottleneckReport, loadDeterminismReport, loadFrontierReport } from './state/reportStore.js'; import { getTopologyForComponent, validateTopologyMap } from './layout/topologyMap.js'; +import { FIRST_RUN_EXAMPLE } from '../shared/first-run-example.js'; const workflow = $('#workflow'); const exportsEl = $('#exports'); @@ -118,6 +119,7 @@ type EntitlementSnapshot = { } | null; const ENTITLEMENT_FRESH_MS = 15 * 60 * 1000; +const FIRST_RUN_COMPLETED_KEY = 'selectpilot_first_run_completed_v1'; type SelectionPreview = { selection: string; @@ -185,6 +187,7 @@ let memorySnapshot: MemoryStatusSnapshot = { lastUpdatedAt: null, }; let entitlementSnapshot: EntitlementSnapshot = null; +let firstRunCompleted = false; const BENCHMARK_CACHE_KEY = 'selectpilot_runtime_benchmark_v1'; const RUNTIME_META_MAX_EVENTS = 6; @@ -916,6 +919,7 @@ async function refreshEntitlementStatus() { entitlementSnapshot = null; } renderEntitlementStatus(); + renderSelectionState(); syncControlAvailability(); } @@ -1002,6 +1006,8 @@ function syncControlAvailability() { intentExecuteButtonEl.disabled = isBusy || !runtimeReady || !selectionReady || !hasIntent; } if (intentClearButtonEl) intentClearButtonEl.disabled = isBusy; + const firstRunButton = document.querySelector('#btn-first-run-example'); + if (firstRunButton) firstRunButton.disabled = isBusy || !runtimeReady || !entitlementSnapshot?.token; } function populatePresetOptions() { @@ -1246,16 +1252,25 @@ function renderSelectionState() { clearNode(selectionCardEl); if (!selectionCardEl) return; + const showFirstRunExample = runtimeSnapshot.ok + && Boolean(entitlementSnapshot?.token) + && !selectionPreview.hasSelection + && !firstRunCompleted; + const header = document.createElement('div'); header.className = 'output-eyebrow'; - header.textContent = selectionPreview.hasSelection ? 'Active selection' : 'No active selection'; + header.textContent = showFirstRunExample + ? 'First result' + : (selectionPreview.hasSelection ? 'Active selection' : 'No active selection'); const title = document.createElement('h3'); - title.textContent = selectionPreview.title || 'Current page'; + title.textContent = showFirstRunExample ? 'See SelectPilot once' : (selectionPreview.title || 'Current page'); const copy = document.createElement('p'); copy.className = 'selection-copy'; - if (selectionPreview.hasSelection) { + if (showFirstRunExample) { + copy.textContent = FIRST_RUN_EXAMPLE.text; + } else if (selectionPreview.hasSelection) { copy.textContent = shorten(selectionPreview.selection, 260); } else if (selectionPreview.pageText) { copy.textContent = 'Extract JSON requires highlighted text. Summarize and Ask can still fall back to page context.'; @@ -1266,9 +1281,20 @@ function renderSelectionState() { const meta = document.createElement('p'); meta.className = 'selection-copy'; const charCount = selectionPreview.hasSelection ? selectionPreview.selection.length : selectionPreview.pageText.length; - meta.textContent = `${selectionPreview.hasSelection ? 'Selection' : 'Page context'} · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; + meta.textContent = showFirstRunExample + ? 'Local example · Action Brief · no page content used' + : `${selectionPreview.hasSelection ? 'Selection' : 'Page context'} · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; selectionCardEl.append(header, title, copy, meta); + if (showFirstRunExample) { + const action = document.createElement('button'); + action.id = 'btn-first-run-example'; + action.className = 'selection-example-action'; + action.type = 'button'; + action.textContent = 'Try local example'; + action.disabled = isBusy; + selectionCardEl.append(action); + } } async function refreshSelectionPreview() { @@ -1377,6 +1403,7 @@ async function refreshRuntime() { setStatus(runtimeSnapshot.error || 'Ollama health check failed'); renderRuntimeState(); } + renderSelectionState(); syncControlAvailability(); } @@ -1428,6 +1455,24 @@ async function doExtract(presetKey?: ExtractionPresetKey) { setStatus('Done'); } +async function doFirstRunExample() { + const selectedPreset = getExtractionPreset(FIRST_RUN_EXAMPLE.preset); + const res = await request('panel:extract_demo'); + renderOutput({ + title: res.label || selectedPreset.label, + markdown: res.markdown || '', + json: res.json || {}, + eyebrow: 'Your first structured result', + meta: 'Created locally from the example above. Ready to read or export.', + exportBase: 'selectpilot-first-result', + }); + renderExports({ markdown: res.markdown, json: res.json, basename: 'selectpilot-first-result' }); + await setJSON(FIRST_RUN_COMPLETED_KEY, true); + firstRunCompleted = true; + renderSelectionState(); + setStatus('Ready to export'); +} + async function doRewrite() { const prompt = agentPromptEl?.value.trim() || 'Rewrite the selected text in clearer, tighter language.'; setStatus('Rewriting...'); @@ -1592,6 +1637,10 @@ function bindActions() { void Promise.all([refreshRuntime(), refreshSelectionPreview(), refreshMemoryStatus(), refreshEntitlementStatus()]); }); $('#btn-extract')?.addEventListener('click', wrap(() => doExtract())); + selectionCardEl?.addEventListener('click', (event) => { + const target = event.target as HTMLElement | null; + if (target?.id === 'btn-first-run-example') void wrap(() => doFirstRunExample())(); + }); $('#btn-summarize')?.addEventListener('click', wrap(() => doSummarize())); $('#btn-rewrite')?.addEventListener('click', wrap(() => doRewrite())); $('#btn-actions')?.addEventListener('click', wrap(() => doActions())); @@ -1693,6 +1742,7 @@ async function initialize() { } setVisiblePanels(['selection_surface', 'runtime_surface', 'report_surface']); + firstRunCompleted = Boolean(await getJSON(FIRST_RUN_COMPLETED_KEY)); renderRuntimeMetaOverlay(); void connectRuntimeMetaStream(); refreshIntentSuggestions(); diff --git a/shared/first-run-example.js b/shared/first-run-example.js new file mode 100644 index 0000000..0c1bfda --- /dev/null +++ b/shared/first-run-example.js @@ -0,0 +1,7 @@ +export const FIRST_RUN_EXAMPLE = Object.freeze({ + version: 'v1', + preset: 'action_brief', + title: 'Launch review notes', + url: 'selectpilot://first-run', + text: 'The launch review is Friday. Maya owns the privacy check by Thursday. Jonas will verify payment and store assets. Publishing stays blocked until both checks pass.', +}); diff --git a/shared/first-run-example.ts b/shared/first-run-example.ts new file mode 100644 index 0000000..fc2107a --- /dev/null +++ b/shared/first-run-example.ts @@ -0,0 +1,17 @@ +import type { ExtractionPresetKey } from '../panel/extraction-presets.js'; + +export type FirstRunExample = Readonly<{ + version: string; + preset: ExtractionPresetKey; + title: string; + url: string; + text: string; +}>; + +export const FIRST_RUN_EXAMPLE: FirstRunExample = Object.freeze({ + version: 'v1', + preset: 'action_brief', + title: 'Launch review notes', + url: 'selectpilot://first-run', + text: 'The launch review is Friday. Maya owns the privacy check by Thursday. Jonas will verify payment and store assets. Publishing stays blocked until both checks pass.', +}); diff --git a/tests/e2e/extension-first-run-flow.spec.mjs b/tests/e2e/extension-first-run-flow.spec.mjs new file mode 100644 index 0000000..7ae7188 --- /dev/null +++ b/tests/e2e/extension-first-run-flow.spec.mjs @@ -0,0 +1,124 @@ +import { test, expect, chromium } from '@playwright/test'; + +test('real extension first-run sends only the canonical local example', async () => { + const executablePath = process.env.SELECTPILOT_CHROME_EXECUTABLE; + test.skip(!executablePath, 'SELECTPILOT_CHROME_EXECUTABLE is required for unpacked-extension proof'); + + const context = await chromium.launchPersistentContext(test.info().outputPath('extension-user-data'), { + executablePath, + headless: process.env.SELECTPILOT_HEADED !== '1', + args: [ + `--disable-extensions-except=${process.cwd()}`, + `--load-extension=${process.cwd()}`, + '--disable-web-security', + ], + }); + + let extractRequest = null; + await context.route('http://127.0.0.1:8083/**', async (route) => { + const request = route.request(); + const path = new URL(request.url()).pathname; + const fulfill = (body) => route.fulfill({ + status: 200, + contentType: 'application/json', + body: JSON.stringify(body), + }); + + if (path === '/profiles') { + return fulfill({ + profiles: [{ + key: 'fast', label: 'Fast', description: 'Local test profile', + generation_model: 'test-generation', embedding_model: 'test-embedding', + target_latency: 'local', intended_for: 'Test', command: 'not-used', + }], + recommended_profile: 'fast', + reason: 'Recommended from local hardware.', + }); + } + if (path === '/health') { + return fulfill({ + ok: true, + ollama: { + reachable: true, model_available: true, active_model: 'test-generation', + privacy_mode: 'local-only', ignored_remote_models: [], + }, + }); + } + if (path === '/privacy-proof') { + return fulfill({ + ok: true, + privacy_mode: 'local-only', + generated_at: new Date().toISOString(), + outbound_observation: { external_calls_registered: false }, + }); + } + if (path === '/benchmark') { + return fulfill({ + ok: true, active_model: 'test-generation', extract_latency_ms: 20, + summarize_latency_ms: 25, recommended_profile: 'fast', + }); + } + if (path === '/runtime-meta/health') { + return fulfill({ ok: false, stream_enabled: false, active_streams: 0, event_version: '1' }); + } + if (path === '/extract') { + extractRequest = request.postDataJSON(); + return fulfill({ + preset: 'action_brief', + label: 'Action Brief', + description: 'Turn selected text into a concise action-oriented brief.', + markdown: '## Action Brief\n\n- Maya: privacy check by Thursday', + json: { + action_items: ['Maya completes the privacy check by Thursday'], + risks: ['Publishing waits for both checks'], + }, + }); + } + return fulfill({ ok: true }); + }); + + try { + let [serviceWorker] = context.serviceWorkers(); + if (!serviceWorker) serviceWorker = await context.waitForEvent('serviceworker'); + const extensionId = new URL(serviceWorker.url()).host; + const panelPage = await context.newPage(); + await panelPage.goto(`chrome-extension://${extensionId}/panel/panel.html`); + await expect(panelPage.locator('#btn-first-run-example')).toHaveCount(0); + + const lockedResponse = await panelPage.evaluate(() => globalThis.chrome.runtime.sendMessage({ type: 'panel:extract_demo' })); + expect(lockedResponse.error).toBe('Paid license required for deterministic extraction'); + expect(extractRequest).toBeNull(); + + await panelPage.evaluate(async () => { + const storage = await import(globalThis.chrome.runtime.getURL('licensing/license-storage.js')); + const now = Date.now(); + await storage.saveLicense({ + token: 'sp_e2e_paid_token', tier: 'essential', issuedAt: now, + expiresAt: now + 86_400_000, cachedAt: now, + }); + }); + await panelPage.reload(); + + await expect(panelPage.locator('#btn-first-run-example')).toBeVisible(); + await panelPage.locator('#btn-first-run-example').click(); + await expect.poll(() => extractRequest).not.toBeNull(); + expect(extractRequest).toEqual({ + text: 'The launch review is Friday. Maya owns the privacy check by Thursday. Jonas will verify payment and store assets. Publishing stays blocked until both checks pass.', + preset: 'action_brief', + url: 'selectpilot://first-run', + title: 'Launch review notes', + metadata: { source: 'selectpilot_first_run', sample_version: 'v1' }, + }); + + await expect(panelPage.locator('#result-title')).toHaveText('Action Brief'); + await expect(panelPage.locator('#exports')).toContainText('Copy Markdown'); + await panelPage.locator('#tab-structured').click(); + await expect(panelPage.locator('#workflow')).toContainText('Publishing waits for both checks'); + await expect(panelPage.locator('#btn-first-run-example')).toHaveCount(0); + + await panelPage.reload(); + await expect(panelPage.locator('#btn-first-run-example')).toHaveCount(0); + } finally { + await context.close(); + } +}); diff --git a/tests/e2e/panel-harness.html b/tests/e2e/panel-harness.html index 89e0bee..906fc77 100644 --- a/tests/e2e/panel-harness.html +++ b/tests/e2e/panel-harness.html @@ -131,7 +131,8 @@

Ready

- \ No newline at end of file + diff --git a/tests/e2e/panel-user-flow.spec.mjs b/tests/e2e/panel-user-flow.spec.mjs index e3abc08..f8eb4c5 100644 --- a/tests/e2e/panel-user-flow.spec.mjs +++ b/tests/e2e/panel-user-flow.spec.mjs @@ -31,4 +31,28 @@ test('sidepanel harness can run extract action and render result', async ({ page await page.click('#btn-sync-order'); await expect(page.locator('#status')).toContainText('Order already synced; entitlement refreshed'); -}); \ No newline at end of file +}); + +test('unlocked first-run uses the dedicated example route and resolves once', async ({ page }) => { + const harnessPath = resolve(process.cwd(), 'tests/e2e/panel-harness.html'); + await page.goto(`${pathToFileURL(harnessPath)}?firstRun=1`); + + await expect(page.locator('#btn-first-run-example')).toHaveCount(0); + await page.fill('#order-id', 'SP-PAID'); + await page.click('#btn-sync-order'); + await expect(page.locator('#btn-first-run-example')).toBeVisible(); + await expect(page.locator('#selection-card')).toContainText('no page content used'); + + await page.click('#btn-first-run-example'); + await expect(page.locator('#result-title')).toHaveText('Action Brief'); + await expect(page.locator('#workflow')).toContainText('Maya'); + await page.click('#tab-structured'); + await expect(page.locator('#workflow')).toContainText('Publishing waits for both checks'); + await expect(page.locator('#exports')).toContainText('Copy Markdown'); + await expect(page.locator('#exports')).toContainText('Copy JSON'); + await expect(page.locator('#btn-first-run-example')).toHaveCount(0); + + const demoMessages = await page.evaluate(() => globalThis.__messageLog.filter((msg) => msg.type === 'panel:extract_demo')); + expect(demoMessages).toHaveLength(1); + expect(demoMessages[0]).toEqual({ type: 'panel:extract_demo' }); +}); diff --git a/tests/panel/first-run-example.test.mjs b/tests/panel/first-run-example.test.mjs new file mode 100644 index 0000000..cb0474e --- /dev/null +++ b/tests/panel/first-run-example.test.mjs @@ -0,0 +1,12 @@ +import test from 'node:test'; +import assert from 'node:assert/strict'; + +import { FIRST_RUN_EXAMPLE } from '../../shared/first-run-example.js'; + +test('first-run example is fixed, local, and uses the existing action preset', () => { + assert.equal(Object.isFrozen(FIRST_RUN_EXAMPLE), true); + assert.equal(FIRST_RUN_EXAMPLE.url, 'selectpilot://first-run'); + assert.equal(FIRST_RUN_EXAMPLE.preset, 'action_brief'); + assert.match(FIRST_RUN_EXAMPLE.text, /privacy check/i); + assert.equal('callerText' in FIRST_RUN_EXAMPLE, false); +}); diff --git a/tests/panel/storage-encryption.test.mjs b/tests/panel/storage-encryption.test.mjs new file mode 100644 index 0000000..a524f85 --- /dev/null +++ b/tests/panel/storage-encryption.test.mjs @@ -0,0 +1,28 @@ +import test from 'node:test'; +import assert from 'node:assert/strict'; + +const values = new Map(); +globalThis.chrome = { + storage: { + local: { + async get(key) { + return { [key]: values.get(key) }; + }, + async set(entries) { + for (const [key, value] of Object.entries(entries)) values.set(key, value); + }, + }, + }, +}; + +const { getDecrypted, setEncrypted } = await import('../../utils/storage.js'); + +test('AES-GCM storage creates a non-exportable per-install key and round-trips data', async () => { + await setEncrypted('test_payload', 'private local value'); + + assert.equal(await getDecrypted('test_payload'), 'private local value'); + assert.equal(Array.isArray(values.get('selectpilot_encryption_key_v1')), true); + assert.equal(values.get('selectpilot_encryption_key_v1').length, 32); + assert.equal(typeof values.get('test_payload'), 'string'); + assert.equal(values.get('test_payload').includes('private local value'), false); +}); diff --git a/utils/storage.js b/utils/storage.js index 94b804d..2284e5e 100644 --- a/utils/storage.js +++ b/utils/storage.js @@ -1,11 +1,24 @@ import { error, log } from './logger.js'; const encoder = new TextEncoder(); const decoder = new TextDecoder(); -const KEY_MATERIAL = 'chromeai_local_key_v1'; -async function getCryptoKey() { - const keyData = encoder.encode(KEY_MATERIAL); +const KEY_STORAGE = 'selectpilot_encryption_key_v1'; +let cryptoKeyPromise = null; +async function loadOrCreateCryptoKey() { + const stored = (await chrome.storage.local.get(KEY_STORAGE))[KEY_STORAGE]; + let keyData; + if (Array.isArray(stored) && stored.length === 32 && stored.every((value) => Number.isInteger(value) && value >= 0 && value <= 255)) { + keyData = new Uint8Array(stored); + } + else { + keyData = crypto.getRandomValues(new Uint8Array(32)); + await chrome.storage.local.set({ [KEY_STORAGE]: Array.from(keyData) }); + } return crypto.subtle.importKey('raw', keyData, { name: 'AES-GCM' }, false, ['encrypt', 'decrypt']); } +function getCryptoKey() { + cryptoKeyPromise ||= loadOrCreateCryptoKey(); + return cryptoKeyPromise; +} export async function setEncrypted(key, value) { try { const cryptoKey = await getCryptoKey(); diff --git a/utils/storage.ts b/utils/storage.ts index 2c2e4d5..4558746 100644 --- a/utils/storage.ts +++ b/utils/storage.ts @@ -2,13 +2,26 @@ import { error, log } from './logger.js'; const encoder = new TextEncoder(); const decoder = new TextDecoder(); -const KEY_MATERIAL = 'chromeai_local_key_v1'; +const KEY_STORAGE = 'selectpilot_encryption_key_v1'; +let cryptoKeyPromise: Promise | null = null; -async function getCryptoKey(): Promise { - const keyData = encoder.encode(KEY_MATERIAL); +async function loadOrCreateCryptoKey(): Promise { + const stored = (await chrome.storage.local.get(KEY_STORAGE))[KEY_STORAGE]; + let keyData: Uint8Array; + if (Array.isArray(stored) && stored.length === 32 && stored.every((value) => Number.isInteger(value) && value >= 0 && value <= 255)) { + keyData = new Uint8Array(stored); + } else { + keyData = crypto.getRandomValues(new Uint8Array(32)); + await chrome.storage.local.set({ [KEY_STORAGE]: Array.from(keyData) }); + } return crypto.subtle.importKey('raw', keyData, { name: 'AES-GCM' }, false, ['encrypt', 'decrypt']); } +function getCryptoKey(): Promise { + cryptoKeyPromise ||= loadOrCreateCryptoKey(); + return cryptoKeyPromise; +} + export async function setEncrypted(key: string, value: string): Promise { try { const cryptoKey = await getCryptoKey(); From e826880698aab48b32e032a8ce7a3bf81d856f0f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Mon, 10 Aug 2026 22:24:04 +0200 Subject: [PATCH 04/66] test(e2e): reuse configured Chrome executable Allow the full browser suite and unpacked-extension proof to run against the same explicitly selected Chrome binary.\n\nVerified: SELECTPILOT_CHROME_EXECUTABLE= pnpm test:e2e (7 passed); pnpm lint --- tests/e2e/playwright.config.mjs | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/tests/e2e/playwright.config.mjs b/tests/e2e/playwright.config.mjs index e9c6f05..4f0f04f 100644 --- a/tests/e2e/playwright.config.mjs +++ b/tests/e2e/playwright.config.mjs @@ -15,6 +15,9 @@ export default defineConfig({ screenshot: 'only-on-failure', launchOptions: { args: ['--allow-file-access-from-files', '--disable-web-security'], + ...(process.env.SELECTPILOT_CHROME_EXECUTABLE + ? { executablePath: process.env.SELECTPILOT_CHROME_EXECUTABLE } + : {}), }, }, -}); \ No newline at end of file +}); From bfab0eb3bff7fd4f1104cca93052e19ba73c3ab9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 11 Aug 2026 09:16:23 +0200 Subject: [PATCH 05/66] fix(server): preserve direct entrypoint startup Restore imports compatible with the documented script invocation and add that invocation as a CI smoke gate. Remove unrelated whitespace noise identified in review.\n\nVerified: python3 server/nano_server.py --help; pnpm lint; pnpm lint:manifest; pnpm typecheck; pnpm test --- .github/workflows/ci.yml | 3 +++ reports/stress/master_summary.json | 1 - server/nano_server.py | 5 ++--- server/ollama_client.py | 2 +- 4 files changed, 6 insertions(+), 5 deletions(-) diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 3c9a3d4..9f0cd61 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -58,5 +58,8 @@ jobs: - name: Build run: pnpm build + - name: Smoke direct server entrypoint + run: python3 server/nano_server.py --help + - name: Test run: pnpm test diff --git a/reports/stress/master_summary.json b/reports/stress/master_summary.json index c826b54..18e1b9a 100644 --- a/reports/stress/master_summary.json +++ b/reports/stress/master_summary.json @@ -64,7 +64,6 @@ "latency_stability": 0.8, "memory_integrity": 0.9, "contract_obedience": 0.95, - "ambiguity_handling": 0.9, "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, diff --git a/server/nano_server.py b/server/nano_server.py index a07551e..26f1d59 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -15,8 +15,8 @@ from datetime import datetime, timezone from urllib.parse import urlparse, parse_qs -from server.ollama_client import OllamaClient, OllamaError -from server.runtime_profiles import build_bootstrap_commands, list_runtime_profiles, recommend_runtime_profile +from ollama_client import OllamaClient, OllamaError +from runtime_profiles import build_bootstrap_commands, list_runtime_profiles, recommend_runtime_profile DEFAULT_PORT = 8083 LOCAL_HOSTS = {"127.0.0.1", "localhost", "::1"} @@ -1807,4 +1807,3 @@ def main(): if __name__ == '__main__': main() - diff --git a/server/ollama_client.py b/server/ollama_client.py index 204dcb0..ffc4acd 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -8,7 +8,7 @@ from urllib.parse import urljoin from urllib.request import Request, urlopen -from server.extraction_presets import get_extraction_preset, render_extraction_markdown +from extraction_presets import get_extraction_preset, render_extraction_markdown def _json_loads_maybe(value: Any) -> Any: From 30a4da7dcc92e4e9b81a149baf2bf0c17c464dbe Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 20:52:00 +0200 Subject: [PATCH 06/66] fix: enforce local runtime contracts --- ...8-18-selectpilot-full-system-completion.md | 144 ++++++++++++++++++ docs/selectpilot-project-snapshot.json | 42 +++++ launchd/com.chromeai.nano.plist | 1 + panel/runtime-profiles.js | 9 +- panel/runtime-profiles.ts | 10 +- scripts/bootstrap-macos-local.sh | 5 +- scripts/install-macos-local.sh | 11 +- scripts/spec-compiler.mjs | 45 +++--- server/nano_server.py | 58 ++++++- server/ollama_client.py | 80 +++++----- server/runtime_profiles.py | 12 +- tests/e2e/extension-first-run-flow.spec.mjs | 3 +- tests/e2e/playwright.config.mjs | 16 ++ tests/server/test_model_feedback.py | 43 ++++++ tests/server/test_ollama_client.py | 63 ++++++++ tests/server/test_runtime_profiles.py | 4 +- 16 files changed, 472 insertions(+), 74 deletions(-) create mode 100644 docs/aegis/plans/2026-08-18-selectpilot-full-system-completion.md create mode 100644 docs/selectpilot-project-snapshot.json create mode 100644 tests/server/test_model_feedback.py create mode 100644 tests/server/test_ollama_client.py diff --git a/docs/aegis/plans/2026-08-18-selectpilot-full-system-completion.md b/docs/aegis/plans/2026-08-18-selectpilot-full-system-completion.md new file mode 100644 index 0000000..a33a1ba --- /dev/null +++ b/docs/aegis/plans/2026-08-18-selectpilot-full-system-completion.md @@ -0,0 +1,144 @@ +# SelectPilot Full-System Completion Plan + +## Aegis Visibility + +This plan prevents slice-level test success, stale generated reports, or an open protected PR from being confused with product completion. The completion owner is the full path from installation and user intent through validated local output, export, protected delivery, and current evidence. + +## Plan Basis + +- User objective: finish SelectPilot to verified full-system Definition of Done without stopping at intermediate slices. +- Product authority: `README.md`, `ROADMAP.md`, `selectpilot_monolith_v3.json`, `ZERO_LEAKAGE.md`, `CONSTITUTION.json`, and active Linear project `SelectPilot — Lokal modell & runtime-stack`. +- Delivery authority: protected GitHub PR #22 and `main`; no self-approval or branch-protection bypass. +- Compute authority: `AGENTS.md` and `docs/compute-distribution.md`. +- Governance authority: signed Axiom lifecycle policy verified by NovaForge at the start of execution. + +## Baseline Usage + +- Required and read: repository contracts above, Linear project and all issues, PR #22 metadata/reviews, NovaForge operator manifest and Axiom lifecycle binding. +- Current delivery candidate: PR #22 head `bfab0eb3bff7fd4f1104cca93052e19ba73c3ab9`. +- Current protected base: `bc2f70bb71b51fbe1e94a4dbe74e427e27e35095`. +- Current contradictions: + - `git diff --check origin/main...HEAD` fails on trailing whitespace/blank-line noise. + - `pnpm test:e2e` does not start the local bridge and fails with `ECONNREFUSED 127.0.0.1:8083`. + - the real unpacked-extension test skips without `SELECTPILOT_CHROME_EXECUTABLE`. + - `reports/verification_report.json` records incomplete verification and `reports/spec_coverage_report.json` records missing/unmapped spec coverage. + - `ROADMAP.md` still lists release-critical onboarding, privacy, E2E, and installation work. + - Linear SOD-617 is correctly externally blocked, while the project-level status update is stale. + +## Requirement Ready Check + +- Goals and scope: explicit from the user and repository/Linear authorities. +- Acceptance: full end-to-end execution, deterministic contract enforcement, no implicit fallback, state integrity, failure/recovery verification, protected `main`, exact merge SHA, CI/E2E/security evidence, and reconciled Linear status. +- Open material authority: independent approval for protected merge remains external; it does not block implementation and verification work. +- Decision: ready. + +## TDD Route + +- Mode: off +- Decision: skipped +- Strict authority: not applicable +- Test posture: diagnostic reproduction plus post-change regression +- Reason: no explicit strict TDD request; existing contract tests and end-to-end gates are the primary evidence surfaces. +- Verification: targeted tests after each repair and the complete remote gate before delivery claims. + +## Scope Fence + +In scope: + +- reconcile current product behavior with repository, Linear, privacy, runtime, and release contracts; +- repair deterministic test/runtime infrastructure; +- complete qualified P0/P1 and product-DoD verticals; +- qualify or explicitly defer roadmap-only model experiments using evidence and dependencies; +- refresh generated evidence so it describes current code truthfully; +- deliver through protected GitHub and update Linear. + +Out of scope without separate authority: + +- production Chrome Web Store activation or paid external publication; +- branch-protection bypass, self-approval, alternate identity, or secret generation; +- cloud inference or telemetry; +- silent local heavy-compute fallback. + +## Compatibility Boundary + +- Preserve the selected-text workflow and existing canonical `/extract` contract. +- Preserve local-only core execution and explicit paid entitlement boundary. +- Preserve deterministic profile/manual override semantics. +- Keep generated JavaScript synchronized from TypeScript; no divergent duplicate behavior. +- Keep documented direct Python entrypoint `python3 server/nano_server.py` working. + +## Change Necessity + +- User-visible need: a genuinely installable, usable, privacy-preserving SelectPilot with reproducible proof. +- Non-code option: documentation-only changes cannot start the bridge, eliminate skipped extension proof, or close runtime/spec gaps. +- Minimum boundary: canonical test configuration and runtime owners first; expand only when a verified product requirement remains unmet. +- Decision: code-change. + +## Architecture Integrity Lens + +- Canonical owners: TypeScript sources for extension behavior, Python bridge for local runtime, Playwright config for E2E lifecycle, spec/report compiler for generated evidence. +- No caller-side fallback may mask missing runtime, model, entitlement, validation, or privacy state. +- Stale reports must be regenerated or retired; they may not remain as false evidence. +- Verdict: repair canonical owners and remove/retire contradictory evidence paths. + +## Execution Readiness View + +- Intent Lock: full-system product completion, not slice completion. +- Scope Fence: product/runtime/release contracts above; no production publication or governance bypass. +- Baseline Lock: PR #22 head plus recorded fresh failures. +- Owner constraints: existing canonical owners; no duplicate runtime or demo engine. +- Test obligations: targeted regression, full remote lint/typecheck/build/unit/E2E/privacy/security, real unpacked-extension path, negative and recovery paths. +- Review gates: independent GitHub approval and protected merge. +- Drift rule: if code, Linear, and product authority conflict, stop the affected edit and reconcile authority before continuing. +- Completion evidence: exact merged SHA on `main`, current CI, runtime/E2E/security proof, current generated reports, Linear evidence, and requirement-by-requirement DoD audit. + +## Task Batches + +### 1. Restore deterministic baseline proof + +- Add a Playwright-owned local bridge lifecycle with model auto-pull disabled and isolated run/log directories. +- Make unpacked-extension proof resolve a supported Chromium executable deterministically and fail explicitly instead of skipping in required proof mode. +- Remove whitespace noise and verify the documented direct Python entrypoint. +- Run remote lint, manifest lint, typecheck, build, unit/server tests, E2E, privacy tests, and `git diff --check`. + +### 2. Reconcile product and generated truth + +- Run the spec compiler, runtime-policy validator, UI-contract validator, stress harness, and report checks against the current candidate. +- Classify every missing spec node and unmapped artifact as implementation drift, obsolete mapping, or explicit roadmap/defer decision. +- Update canonical mappings/implementation or retire obsolete generated evidence; do not edit reports by hand to manufacture green status. + +### 3. Complete release-critical user verticals + +- Verify one-command bootstrap including Ollama detection, model provisioning, launchd/server health, explicit failure, and final summary. +- Verify real selected-text → side panel → local request → validated result → export flow. +- Enforce full-extension no-external-network regression. +- Reconcile first-run, copy, status terminology, controls, and accessibility with MUE and product truth. + +### 4. Complete qualified runtime/model verticals + +- Re-evaluate SOD-395, SOD-396, and SOD-394 against current code and actual supported Ollama models. +- Implement explicit per-profile `num_ctx` and JSON-schema constrained structured output where not already complete. +- Change default model only with current availability/license/runtime evidence and hardware-envelope proof; otherwise correct the Linear/project contract instead of shipping a false default. +- Keep SOD-397–SOD-401 roadmap-only unless their stated activation conditions are proven. + +### 5. Close security, privacy, reliability, and release gaps + +- Run dependency/security audits and secret scans on the exact candidate. +- Verify failure/retry, entitlement, encrypted persistence, direct entrypoint, no implicit model fallback, and no raw-selection/prompt leakage. +- Reconcile store assets, privacy policy, support/homepage requirements, packaging, and release checklist without claiming external publication. +- Perform gap, blind-spot, and negative-space analysis; close or explicitly block every material finding. + +### 6. Governed delivery and final audit + +- Push coherent verified commits to PR #22's branch so auto-merge remains active. +- Preserve SOD-617 as `In Review` + `Externally Blocked` until independent approval and protected merge occur. +- After approval, verify auto-merge, exact merge SHA/ancestry on `main`, full CI/E2E/security evidence, and post-merge runtime smoke. +- Update SOD-617 and project-level Linear status/evidence consistently. +- Mark the product complete only after every plan requirement has current authoritative proof. + +## Risks and Retirement + +- Risk: the monolith spec contains stale path mappings. Retirement requires compiler-backed replacement mappings and removal of contradictory generated reports. +- Risk: Linux Chromium proof may differ from macOS installation/runtime proof. Final evidence must include both remote automated proof and a bounded Mac control-surface smoke where platform-specific behavior is required. +- Risk: PR #22 independent approval remains external. Implementation continues; delivery status remains blocked until the protected merge is verified. +- Compatibility helpers or temporary test paths must include an explicit retirement trigger and may not become silent fallback behavior. diff --git a/docs/selectpilot-project-snapshot.json b/docs/selectpilot-project-snapshot.json new file mode 100644 index 0000000..b518af0 --- /dev/null +++ b/docs/selectpilot-project-snapshot.json @@ -0,0 +1,42 @@ +{ + "schema_version": "1.0.0", + "project_key": "/home/krille/workspaces/selectpilot", + "repository": "kristoffersodersten/selectpilot", + "captured_at": "2026-08-18T18:45:00Z", + "delivery": { + "branch": "codex/ci-foundation-repair", + "head_sha": "bfab0eb3bff7fd4f1104cca93052e19ba73c3ab9", + "protected_base": "main", + "base_sha": "bc2f70bb71b51fbe1e94a4dbe74e427e27e35095", + "pull_request": 22, + "auto_merge": true, + "independent_approval_observed": false + }, + "linear": { + "project": "SelectPilot — Lokal modell & runtime-stack", + "project_id": "ff713009-3fb7-4334-915d-12a1f732b543", + "active_issue": "SOD-617", + "active_issue_state": "In Review", + "active_issue_labels": ["Externally Blocked"] + }, + "governance": { + "axiom_policy_result": "AXIOM_LIFECYCLE_POLICY_VERIFIED", + "axiom_policy_version": "1.0.0", + "axiom_policy_sha256": "932bbb231cb5e1aeb35d6bbe6327db6f3b1b3cf1e125e38366cbccbbc59434cf", + "axiom_public_key_sha256": "179b5520ccd5f32f04f0fc09df59a619975a2f1c93b24674883b07fa0ce301b8", + "novaforge_source_sha": "757e8e3151a96e88ff67b0aba78c6f785181ae90", + "operator_verification": "NOVAFORGE_OPERATOR_VERIFIED", + "remote_compute_audit": "PASS" + }, + "fresh_baseline": { + "lint": "PASS", + "manifest_lint": "PASS", + "typecheck": "PASS", + "build": "PASS", + "unit_and_server_tests": "PASS: 5 panel, 25 server", + "e2e": "FAIL: 2 ECONNREFUSED, 4 pass, 1 skip", + "diff_check": "FAIL: server/runtime_profiles.py blank line at EOF" + }, + "next_action": "Repair deterministic Playwright bridge lifecycle and required unpacked-extension proof, then rerun the complete remote baseline.", + "completion_state": "NOT_COMPLETE" +} diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index 8ef8571..56b69c5 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -25,6 +25,7 @@ CHROMEAI_OLLAMA_BASE_URL__OLLAMA_BASE_URL__ CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ CHROMEAI_OLLAMA_EMBED_MODEL__OLLAMA_EMBED_MODEL__ + CHROMEAI_OLLAMA_NUM_CTX__OLLAMA_NUM_CTX__ CHROMEAI_BINARY_HASH__BINARY_HASH__ diff --git a/panel/runtime-profiles.js b/panel/runtime-profiles.js index 0d7c9ab..8bba413 100644 --- a/panel/runtime-profiles.js +++ b/panel/runtime-profiles.js @@ -3,8 +3,9 @@ export const RUNTIME_PROFILES = [ key: 'fast', label: 'Fast', description: 'Smallest viable local profile for structured extraction and low-latency summaries.', - generation_model: 'qwen2.5:0.5b', + generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', + num_ctx: 16384, target_latency: '1-4s', intended_for: 'Selected-text extraction, action briefs, and quick summaries.', command: './scripts/bootstrap-macos-local.sh --profile fast', @@ -14,8 +15,9 @@ export const RUNTIME_PROFILES = [ key: 'balanced', label: 'Balanced', description: 'Higher quality local profile for rewrite and general-purpose browser transforms.', - generation_model: 'qwen2.5:3b', + generation_model: 'gemma4:e4b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', + num_ctx: 32768, target_latency: '2-6s', intended_for: 'Daily use when you want better quality without drifting into heavy models.', command: './scripts/bootstrap-macos-local.sh --profile balanced', @@ -24,8 +26,9 @@ export const RUNTIME_PROFILES = [ key: 'advanced', label: 'Advanced', description: 'Manual opt-in profile for stronger reasoning on larger machines.', - generation_model: 'qwen2.5:7b', + generation_model: 'gemma4:12b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', + num_ctx: 32768, target_latency: '4-10s', intended_for: 'Heavier rewrite and ask flows when latency budget is less important.', command: './scripts/bootstrap-macos-local.sh --profile advanced', diff --git a/panel/runtime-profiles.ts b/panel/runtime-profiles.ts index 90beab6..228d721 100644 --- a/panel/runtime-profiles.ts +++ b/panel/runtime-profiles.ts @@ -6,6 +6,7 @@ export type RuntimeProfile = { description: string; generation_model: string; embedding_model: string; + num_ctx: number; target_latency: string; intended_for: string; command: string; @@ -17,8 +18,9 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ key: 'fast', label: 'Fast', description: 'Smallest viable local profile for structured extraction and low-latency summaries.', - generation_model: 'qwen2.5:0.5b', + generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', + num_ctx: 16384, target_latency: '1-4s', intended_for: 'Selected-text extraction, action briefs, and quick summaries.', command: './scripts/bootstrap-macos-local.sh --profile fast', @@ -28,8 +30,9 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ key: 'balanced', label: 'Balanced', description: 'Higher quality local profile for rewrite and general-purpose browser transforms.', - generation_model: 'qwen2.5:3b', + generation_model: 'gemma4:e4b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', + num_ctx: 32768, target_latency: '2-6s', intended_for: 'Daily use when you want better quality without drifting into heavy models.', command: './scripts/bootstrap-macos-local.sh --profile balanced', @@ -38,8 +41,9 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ key: 'advanced', label: 'Advanced', description: 'Manual opt-in profile for stronger reasoning on larger machines.', - generation_model: 'qwen2.5:7b', + generation_model: 'gemma4:12b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', + num_ctx: 32768, target_latency: '4-10s', intended_for: 'Heavier rewrite and ask flows when latency budget is less important.', command: './scripts/bootstrap-macos-local.sh --profile advanced', diff --git a/scripts/bootstrap-macos-local.sh b/scripts/bootstrap-macos-local.sh index 6466cb7..31f17d0 100755 --- a/scripts/bootstrap-macos-local.sh +++ b/scripts/bootstrap-macos-local.sh @@ -62,6 +62,7 @@ print(json.dumps({ "reason": recommendation["reason"], "generation_model": runtime_profile.generation_model, "embedding_model": runtime_profile.embedding_model, + "num_ctx": runtime_profile.num_ctx, "command": commands["command"], })) PY @@ -74,7 +75,7 @@ import shlex import sys payload = json.loads(sys.argv[1]) -for key in ("selected_profile", "generation_model", "embedding_model", "reason"): +for key in ("selected_profile", "generation_model", "embedding_model", "num_ctx", "reason"): print(f"{key.upper()}={shlex.quote(str(payload[key]))}") PY )" @@ -123,6 +124,7 @@ fi CHROMEAI_OLLAMA_MODEL="$GEN_MODEL" \ CHROMEAI_OLLAMA_EMBED_MODEL="$EMBED_MODEL" \ +CHROMEAI_OLLAMA_NUM_CTX="$NUM_CTX" \ "$ROOT/scripts/install-macos-local.sh" STATUS_LAUNCHAGENT="ok" @@ -140,6 +142,7 @@ Profile: $SELECTED_PROFILE Reason: $REASON Generation model: $GEN_MODEL Embedding model: $EMBED_MODEL +Context window: $NUM_CTX Next recommended command: pnpm benchmark:local diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index 3e74fbb..c10c2f1 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -6,8 +6,9 @@ TEMPLATE="$ROOT/launchd/com.chromeai.nano.plist" DEST="${HOME}/Library/LaunchAgents/com.chromeai.nano.plist" HASH="$(shasum -a 256 "$ROOT/server/nano_server.py" | awk '{print $1}')" OLLAMA_BASE_URL="${CHROMEAI_OLLAMA_BASE_URL:-http://127.0.0.1:11434}" -OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-qwen2.5:0.5b}" +OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-gemma4:e2b-it-qat}" OLLAMA_EMBED_MODEL="${CHROMEAI_OLLAMA_EMBED_MODEL:-nomic-embed-text-v2-moe:latest}" +OLLAMA_NUM_CTX="${CHROMEAI_OLLAMA_NUM_CTX:-16384}" RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run}" LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}" @@ -22,6 +23,7 @@ sed \ -e "s|__OLLAMA_BASE_URL__|$OLLAMA_BASE_URL|g" \ -e "s|__OLLAMA_MODEL__|$OLLAMA_MODEL|g" \ -e "s|__OLLAMA_EMBED_MODEL__|$OLLAMA_EMBED_MODEL|g" \ + -e "s|__OLLAMA_NUM_CTX__|$OLLAMA_NUM_CTX|g" \ "$TEMPLATE" > "$DEST" launchctl unload "$DEST" 2>/dev/null || true @@ -37,7 +39,8 @@ Next steps: 2. Current Ollama base URL: $OLLAMA_BASE_URL 3. Current Ollama model: $OLLAMA_MODEL 4. Local bridge URL: http://127.0.0.1:8083 - 5. Run dir: $RUN_DIR - 6. Log dir: $LOG_DIR - 7. Run 'pnpm benchmark:local' to validate latency on this machine. + 5. Ollama context window: $OLLAMA_NUM_CTX + 6. Run dir: $RUN_DIR + 7. Log dir: $LOG_DIR + 8. Run 'pnpm benchmark:local' to validate latency on this machine. EOF diff --git a/scripts/spec-compiler.mjs b/scripts/spec-compiler.mjs index 32c990f..a0010a3 100644 --- a/scripts/spec-compiler.mjs +++ b/scripts/spec-compiler.mjs @@ -14,6 +14,7 @@ const SPEC_PATH = path.resolve(repoRoot, 'selectpilot_monolith_v3.json'); const BENCHMARK_SPEC_PATH = path.resolve(repoRoot, 'selectpilot_benchmark_v1.json'); const REPORTS_DIR = path.resolve(repoRoot, 'reports'); const SHARED_TYPES_DIR = path.resolve(repoRoot, 'shared', 'types'); +const REQUIRED_SHARED_TYPE_FILES = ['runtime.ts', 'model.ts', 'intent.ts', 'events.ts']; const VALID_MODES = new Set(['full', 'single', 'family', 'frontier', 'determinism', 'engine']); @@ -419,16 +420,20 @@ function buildPhaseReports(spec, moduleManifest, missingTopNodes) { })); } -function buildSharedTypes(spec) { - const taskTypes = ['extract', 'summarize', 'agent']; - const candidateModels = spec.simulation_and_benchmarking?.candidate_models || []; - - return { - 'runtime.ts': `export type RuntimeStatus = 'idle' | 'running' | 'completed' | 'error';\n`, - 'model.ts': `export type TaskType = ${taskTypes.map((v) => `'${v}'`).join(' | ')};\nexport type CandidateModelId = ${candidateModels.length ? candidateModels.map((v) => `'${v}'`).join(' | ') : 'string'};\n`, - 'intent.ts': `export type CompiledIntent = { clarify_required: boolean; ambiguity_score: number; };\n`, - 'events.ts': `export type TerminalEventType = 'RUNTIME_COMPLETED' | 'RUNTIME_FAILED';\n`, - }; +async function validateSharedTypes() { + const missing = []; + for (const fileName of REQUIRED_SHARED_TYPE_FILES) { + const filePath = path.join(SHARED_TYPES_DIR, fileName); + const source = await fsp.readFile(filePath, 'utf8').catch(() => ''); + if (!source.trim() || !/\bexport\s+(?:type|interface)\b/.test(source)) { + missing.push(path.relative(repoRoot, filePath)); + } + } + if (missing.length) { + fail('typed_contract_missing', 'Required maintained shared type contracts are missing or empty', { + missing, + }); + } } function buildVerificationReport(spec, traceability, requiredRefs) { @@ -714,7 +719,7 @@ async function main() { } await ensureDir(REPORTS_DIR); - await ensureDir(SHARED_TYPES_DIR); + await validateSharedTypes(); const requiredRefs = collectRequiredSpecRefs(spec); const traceability = await traceabilityScan(); @@ -728,11 +733,6 @@ async function main() { const architectureDecisionReport = buildArchitectureDecisionReport(spec); const benchmarkOutputs = buildBenchmarkOutputs(benchmarkSpec, mode); - const sharedTypes = buildSharedTypes(spec); - for (const [fileName, source] of Object.entries(sharedTypes)) { - await writeText(path.join(SHARED_TYPES_DIR, fileName), source); - } - await writeJson(path.join(REPORTS_DIR, 'repo_plan.json'), repoPlan); await writeJson(path.join(REPORTS_DIR, 'module_manifest.json'), moduleManifest); await writeJson(path.join(REPORTS_DIR, 'phase_report.json'), phaseReports); @@ -752,8 +752,8 @@ async function main() { await writeJson(path.join(REPORTS_DIR, 'engine_decision_report.json'), benchmarkOutputs.engineDecision); await writeText(path.join(REPORTS_DIR, 'frontier_summary.html'), buildFrontierSummaryHtml(benchmarkOutputs.frontierDecisions)); - console.log(stableStringify({ - ok: true, + const result = { + ok: verificationReport.pass_fail_summary.ok, mode, input: path.relative(repoRoot, SPEC_PATH), benchmark_input: path.relative(repoRoot, BENCHMARK_SPEC_PATH), @@ -777,7 +777,14 @@ async function main() { 'reports/engine_decision_report.json', 'reports/frontier_summary.html', ], - })); + }; + if (!result.ok) { + fail('spec_coverage_incomplete', 'Specification coverage is incomplete', { + reports: result.reports, + missing_spec_refs: verificationReport.pass_fail_summary.missing_spec_refs, + }); + } + console.log(stableStringify(result)); } main().catch((err) => { diff --git a/server/nano_server.py b/server/nano_server.py index 26f1d59..0def16e 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -536,19 +536,67 @@ def _now_ms() -> int: return int(time.time() * 1000) +def _recent_model_feedback(model_id: str, limit: int = MODEL_FAILURE_ISOLATION_WINDOW) -> list[dict]: + if not model_id or not LIVE_FEEDBACK_PATH.exists(): + return [] + records: list[dict] = [] + try: + with _LIVE_FEEDBACK_LOCK: + lines = LIVE_FEEDBACK_PATH.read_text(encoding="utf-8").splitlines() + for line in reversed(lines): + try: + item = json.loads(line) + except json.JSONDecodeError: + continue + if item.get("type") == "model_feedback" and item.get("model_id") == model_id: + records.append(item) + if len(records) >= limit: + break + except OSError: + return [] + return list(reversed(records)) + + def record_model_feedback(model_id: str, *, success: bool, retries: int, latency_ms: int, cancelled: bool = False) -> None: - return + if not model_id: + return + _append_live_feedback({ + "type": "model_feedback", + "model_id": model_id, + "success": bool(success), + "retries": max(0, int(retries)), + "latency_ms": max(0, int(latency_ms)), + "cancelled": bool(cancelled), + }) def _is_quarantined(model_id: str) -> bool: - return False + records = _recent_model_feedback(model_id) + if len(records) < MODEL_FAILURE_ISOLATION_WINDOW: + return False + failures = sum(1 for item in records if not item.get("success") or item.get("cancelled")) + return (failures / len(records)) > MODEL_FAILURE_ISOLATION_THRESHOLD def _recent_feedback_penalty(model_id: str) -> float: - return 0.0 + records = _recent_model_feedback(model_id) + if not records: + return 0.0 + failure_ratio = sum(1 for item in records if not item.get("success") or item.get("cancelled")) / len(records) + retry_ratio = sum(min(3, int(item.get("retries") or 0)) for item in records) / (len(records) * 3) + latency_overrun_ratio = sum(1 for item in records if int(item.get("latency_ms") or 0) > 10_000) / len(records) + return round(failure_ratio + (retry_ratio * 0.25) + (latency_overrun_ratio * 0.1), 4) def apply_hysteresis(task_type: str, selected_model_id: str, available_model_ids: list[str]) -> str: + del task_type + if selected_model_id not in available_model_ids or _is_quarantined(selected_model_id): + raise ValidationError( + "selected_model_unavailable", + "Selected model is unavailable or quarantined", + status=503, + details={"selected_model_id": selected_model_id}, + ) return selected_model_id @@ -975,6 +1023,8 @@ def select_smallest_sufficient_model(task_analysis: dict) -> dict: }, ) + selected_id = apply_hysteresis(task_type, selected_id, available_ids) + selected = next(item[0] for item in sorted_candidates if str(item[0].get("id") or "") == selected_id) selected_penalty = _recent_feedback_penalty(selected_id) return { "model": selected.get("id"), @@ -1656,6 +1706,8 @@ def _execute_summarize(): return except (RuntimeError, OllamaError) as e: duration_ms = round((time.perf_counter() - started_at) * 1000) + if runtime_model_id and path in {'/summarize', '/agent', '/extract'}: + record_model_feedback(runtime_model_id, success=False, retries=retry_count, latency_ms=duration_ms) _append_live_feedback({ "trace_id": trace_id, "operation": contract.name, diff --git a/server/ollama_client.py b/server/ollama_client.py index ffc4acd..60e62f4 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -94,6 +94,7 @@ class OllamaConfig: model: str embed_model: str timeout_seconds: float + num_ctx: int class OllamaError(RuntimeError): @@ -105,6 +106,7 @@ def __init__(self, config: OllamaConfig | None = None): if config is None: default_generation_model = "llama3.2" default_embed_model = "nomic-embed-text-v2-moe:latest" + default_num_ctx = 16384 runtime_profile = os.environ.get("CHROMEAI_RUNTIME_PROFILE", "auto") try: @@ -115,6 +117,7 @@ def __init__(self, config: OllamaConfig | None = None): profile = get_runtime_profile(resolved_profile) default_generation_model = profile.generation_model default_embed_model = profile.embedding_model + default_num_ctx = profile.num_ctx except Exception: pass @@ -123,6 +126,7 @@ def __init__(self, config: OllamaConfig | None = None): model=os.environ.get("CHROMEAI_OLLAMA_MODEL", default_generation_model), embed_model=os.environ.get("CHROMEAI_OLLAMA_EMBED_MODEL", default_embed_model), timeout_seconds=float(os.environ.get("CHROMEAI_OLLAMA_TIMEOUT_SECONDS", "30")), + num_ctx=int(os.environ.get("CHROMEAI_OLLAMA_NUM_CTX", str(default_num_ctx))), ) self.config = config @@ -210,16 +214,41 @@ def _model_names(self, local_only: bool = False) -> list[str]: return models def _resolve_model(self, requested: str, preferences: list[str], models: list[str]) -> str: - if requested in models: - return requested - for candidate in preferences: - for model in models: - if model == candidate or model.startswith(f"{candidate}:"): - return model - if models: - return models[0] + del preferences + if self._model_available_locally(requested, models): + if requested in models: + return requested + requested_base = requested.split(":", 1)[0] + return next(model for model in models if model == requested_base or model.startswith(f"{requested_base}:")) return requested + def _generation_options(self, temperature: float) -> dict[str, Any]: + return {"temperature": temperature, "num_ctx": self.config.num_ctx} + + def _require_structured_object(self, raw_response: str, schema: dict[str, Any], operation: str) -> dict[str, Any]: + content = _parse_jsonish(raw_response) + if not isinstance(content, dict): + raise OllamaError(f"{operation} returned non-object structured output") + + missing = [key for key in schema.get("required", []) if key not in content] + if missing: + raise OllamaError(f"{operation} structured output is missing required fields: {', '.join(missing)}") + + properties = schema.get("properties", {}) + if schema.get("additionalProperties") is False: + unexpected = sorted(set(content) - set(properties)) + if unexpected: + raise OllamaError(f"{operation} structured output has unexpected fields: {', '.join(unexpected)}") + + type_map = {"string": str, "array": list, "object": dict} + for key, definition in properties.items(): + if key not in content: + continue + expected = type_map.get(definition.get("type")) + if expected is not None and not isinstance(content[key], expected): + raise OllamaError(f"{operation} structured output field {key} has invalid type") + return content + def _request_json(self, path: str, payload: dict[str, Any] | None = None) -> Any: url = urljoin(self.config.base_url + "/", path.lstrip("/")) data = None if payload is None else json.dumps(payload).encode("utf-8") @@ -263,6 +292,7 @@ def health(self) -> dict[str, Any]: "active_model": self.config.model, "active_embed_model": self.config.embed_model, "timeout_seconds": self.config.timeout_seconds, + "num_ctx": self.config.num_ctx, "reachable": False, "status": "degraded", "error": str(e), @@ -286,6 +316,7 @@ def health(self) -> dict[str, Any]: "active_model": active_model, "active_embed_model": active_embed_model, "timeout_seconds": self.config.timeout_seconds, + "num_ctx": self.config.num_ctx, } return { **base, @@ -338,19 +369,11 @@ def summarize(self, text: str, title: str | None = None, url: str | None = None, "system": "You write precise summaries for selected text in a browser side panel.", "stream": False, "format": schema, - "options": {"temperature": 0.2}, + "options": self._generation_options(0.2), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() - content = _parse_jsonish(raw_response) - if not isinstance(content, dict): - content = { - "summary": raw_response or "No summary produced.", - "bullets": [], - "action_items": [], - "title": title or "Summary", - "tags": ["ollama", "fallback"], - } + content = self._require_structured_object(raw_response, schema, "summarize") summary = str(content.get("summary", "")).strip() bullets = [str(item).strip() for item in content.get("bullets", []) if str(item).strip()] @@ -401,20 +424,11 @@ def agent(self, prompt: str, context: dict[str, Any] | None = None) -> dict[str, "system": "You are a practical browser copilot that rewrites and structures selected text locally.", "stream": False, "format": schema, - "options": {"temperature": 0.2}, + "options": self._generation_options(0.2), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() - content = _parse_jsonish(raw_response) - if not isinstance(content, dict): - content = { - "reasoning": ["Model response was not valid JSON"], - "markdown": raw_response or "No response produced.", - "json": { - "raw": raw_response, - "prompt": prompt, - }, - } + content = self._require_structured_object(raw_response, schema, "agent") if "reasoning" not in content or "markdown" not in content or "json" not in content: result = content.get("result", {}) if isinstance(content.get("result"), dict) else {} @@ -509,13 +523,11 @@ def extract( "system": "You generate clean structured extraction results for highlighted browser text.", "stream": False, "format": preset.schema, - "options": {"temperature": 0.1}, + "options": self._generation_options(0.1), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() - content = _parse_jsonish(raw_response) - if not isinstance(content, dict): - content = {} + content = self._require_structured_object(raw_response, preset.schema, "extract") normalized: dict[str, Any] = {} schema_props = preset.schema.get("properties", {}) @@ -532,7 +544,7 @@ def extract( normalized[key] = str(raw_value or "").strip() if not any(normalized.values()): - normalized[preset.intro_key] = raw_response or "No extraction produced." + raise OllamaError("extract structured output contains no usable values") markdown = render_extraction_markdown(preset, normalized) return { diff --git a/server/runtime_profiles.py b/server/runtime_profiles.py index 8080709..97496cf 100644 --- a/server/runtime_profiles.py +++ b/server/runtime_profiles.py @@ -15,6 +15,7 @@ class RuntimeProfile: description: str generation_model: str embedding_model: str + num_ctx: int target_latency: str intended_for: str is_default_auto: bool = False @@ -25,8 +26,9 @@ class RuntimeProfile: key="fast", label="Fast", description="Smallest viable local profile for structured extraction and low-latency summaries.", - generation_model="qwen2.5:0.5b", + generation_model="gemma4:e2b-it-qat", embedding_model="nomic-embed-text-v2-moe:latest", + num_ctx=16384, target_latency="1-4s", intended_for="Selected-text extraction, action briefs, and quick summaries.", is_default_auto=True, @@ -35,8 +37,9 @@ class RuntimeProfile: key="balanced", label="Balanced", description="Higher quality local profile for rewrite and general-purpose browser transforms.", - generation_model="qwen2.5:3b", + generation_model="gemma4:e4b-it-qat", embedding_model="nomic-embed-text-v2-moe:latest", + num_ctx=32768, target_latency="2-6s", intended_for="Daily use when you want better quality without drifting into heavy models.", ), @@ -44,8 +47,9 @@ class RuntimeProfile: key="advanced", label="Advanced", description="Manual opt-in profile for stronger reasoning on larger machines.", - generation_model="qwen2.5:7b", + generation_model="gemma4:12b-it-qat", embedding_model="nomic-embed-text-v2-moe:latest", + num_ctx=32768, target_latency="4-10s", intended_for="Heavier rewrite and ask flows when latency budget is less important.", ), @@ -131,5 +135,5 @@ def build_bootstrap_commands(profile_key: str, project_root: str | Path) -> dict "command": command, "generation_model": profile.generation_model, "embedding_model": profile.embedding_model, + "num_ctx": profile.num_ctx, } - diff --git a/tests/e2e/extension-first-run-flow.spec.mjs b/tests/e2e/extension-first-run-flow.spec.mjs index 7ae7188..05a2a8b 100644 --- a/tests/e2e/extension-first-run-flow.spec.mjs +++ b/tests/e2e/extension-first-run-flow.spec.mjs @@ -1,8 +1,7 @@ import { test, expect, chromium } from '@playwright/test'; test('real extension first-run sends only the canonical local example', async () => { - const executablePath = process.env.SELECTPILOT_CHROME_EXECUTABLE; - test.skip(!executablePath, 'SELECTPILOT_CHROME_EXECUTABLE is required for unpacked-extension proof'); + const executablePath = process.env.SELECTPILOT_CHROME_EXECUTABLE || chromium.executablePath(); const context = await chromium.launchPersistentContext(test.info().outputPath('extension-user-data'), { executablePath, diff --git a/tests/e2e/playwright.config.mjs b/tests/e2e/playwright.config.mjs index 4f0f04f..85803d8 100644 --- a/tests/e2e/playwright.config.mjs +++ b/tests/e2e/playwright.config.mjs @@ -1,4 +1,7 @@ import { defineConfig } from '@playwright/test'; +import { fileURLToPath } from 'node:url'; + +const projectRoot = fileURLToPath(new URL('../..', import.meta.url)); export default defineConfig({ testDir: './', @@ -8,6 +11,19 @@ export default defineConfig({ retries: process.env.CI ? 2 : 0, workers: process.env.CI ? 1 : undefined, reporter: [['list']], + webServer: { + command: 'python3 server/nano_server.py --run-dir /tmp/selectpilot-e2e-run --log-dir /tmp/selectpilot-e2e-log', + cwd: projectRoot, + env: { + ...process.env, + CHROMEAI_AUTO_PULL_MODELS: '0', + }, + url: 'http://127.0.0.1:8083/health', + reuseExistingServer: false, + timeout: 30_000, + stdout: 'pipe', + stderr: 'pipe', + }, use: { baseURL: 'http://127.0.0.1:8083', headless: true, diff --git a/tests/server/test_model_feedback.py b/tests/server/test_model_feedback.py new file mode 100644 index 0000000..cb1e7d4 --- /dev/null +++ b/tests/server/test_model_feedback.py @@ -0,0 +1,43 @@ +from __future__ import annotations + +import json +import sys +import tempfile +import unittest +from pathlib import Path +from unittest.mock import patch + +ROOT = Path(__file__).resolve().parents[2] +SERVER_DIR = ROOT / "server" +if str(SERVER_DIR) not in sys.path: + sys.path.insert(0, str(SERVER_DIR)) + +import nano_server # noqa: E402 + + +class ModelFeedbackTests(unittest.TestCase): + def test_failure_window_quarantines_model(self) -> None: + with tempfile.TemporaryDirectory() as temp_dir: + feedback_path = Path(temp_dir) / "feedback.jsonl" + rows = [ + {"type": "model_feedback", "model_id": "model:a", "success": index >= 3, "retries": 0, "latency_ms": 10} + for index in range(10) + ] + feedback_path.write_text("\n".join(json.dumps(row) for row in rows) + "\n", encoding="utf-8") + with patch.object(nano_server, "LIVE_FEEDBACK_PATH", feedback_path): + self.assertTrue(nano_server._is_quarantined("model:a")) + self.assertFalse(nano_server._is_quarantined("model:b")) + + def test_penalty_reflects_retries_and_latency(self) -> None: + with tempfile.TemporaryDirectory() as temp_dir: + feedback_path = Path(temp_dir) / "feedback.jsonl" + feedback_path.write_text( + json.dumps({"type": "model_feedback", "model_id": "model:a", "success": True, "retries": 3, "latency_ms": 12000}) + "\n", + encoding="utf-8", + ) + with patch.object(nano_server, "LIVE_FEEDBACK_PATH", feedback_path): + self.assertEqual(nano_server._recent_feedback_penalty("model:a"), 0.35) + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/server/test_ollama_client.py b/tests/server/test_ollama_client.py new file mode 100644 index 0000000..b49c990 --- /dev/null +++ b/tests/server/test_ollama_client.py @@ -0,0 +1,63 @@ +from __future__ import annotations + +import sys +import unittest +from pathlib import Path + +ROOT = Path(__file__).resolve().parents[2] +SERVER_DIR = ROOT / "server" +if str(SERVER_DIR) not in sys.path: + sys.path.insert(0, str(SERVER_DIR)) + +from ollama_client import OllamaClient, OllamaConfig, OllamaError # noqa: E402 + + +class StubOllamaClient(OllamaClient): + def __init__(self, response: dict): + super().__init__(OllamaConfig( + base_url="http://127.0.0.1:11434", + model="gemma4:e2b-it-qat", + embed_model="nomic-embed-text-v2-moe:latest", + timeout_seconds=1, + num_ctx=16384, + )) + self.response = response + self.last_payload = None + + def _model_names(self, local_only: bool = False) -> list[str]: + return [self.config.model, self.config.embed_model] + + def _request_json(self, path: str, payload=None): + self.last_payload = payload + return self.response + + +class OllamaClientContractTests(unittest.TestCase): + def test_explicit_model_is_not_silently_substituted(self) -> None: + client = StubOllamaClient({}) + self.assertEqual( + client.active_generation_model(["qwen2.5:0.5b"]), + "gemma4:e2b-it-qat", + ) + + def test_summarize_sends_explicit_context_window(self) -> None: + client = StubOllamaClient({ + "model": "gemma4:e2b-it-qat", + "response": '{"summary":"S","bullets":[],"action_items":[],"title":"T","tags":[]}', + }) + client.summarize("content") + self.assertEqual(client.last_payload["options"]["num_ctx"], 16384) + + def test_summarize_rejects_non_schema_output(self) -> None: + client = StubOllamaClient({"response": "not json"}) + with self.assertRaisesRegex(OllamaError, "non-object structured output"): + client.summarize("content") + + def test_summarize_rejects_missing_required_fields(self) -> None: + client = StubOllamaClient({"response": '{"summary":"S"}'}) + with self.assertRaisesRegex(OllamaError, "missing required fields"): + client.summarize("content") + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/server/test_runtime_profiles.py b/tests/server/test_runtime_profiles.py index 114b317..59eb57a 100644 --- a/tests/server/test_runtime_profiles.py +++ b/tests/server/test_runtime_profiles.py @@ -16,7 +16,8 @@ class RuntimeProfileTests(unittest.TestCase): def test_unknown_profile_falls_back_to_fast(self) -> None: profile = get_runtime_profile("missing") self.assertEqual(profile.key, "fast") - self.assertEqual(profile.generation_model, "qwen2.5:0.5b") + self.assertEqual(profile.generation_model, "gemma4:e2b-it-qat") + self.assertEqual(profile.num_ctx, 16384) def test_auto_recommendation_prefers_balanced_on_large_machines(self) -> None: recommendation = recommend_runtime_profile( @@ -27,6 +28,7 @@ def test_auto_recommendation_prefers_balanced_on_large_machines(self) -> None: def test_bootstrap_command_contains_profile(self) -> None: command = build_bootstrap_commands("balanced", ROOT) self.assertIn("--profile balanced", command["command"]) + self.assertEqual(command["num_ctx"], 32768) if __name__ == "__main__": From e8cf5af1a02e3478b516c00095c105731aefe79f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 20:53:28 +0200 Subject: [PATCH 07/66] fix(deps): patch brace expansion advisories --- pnpm-lock.yaml | 13 ++++++++----- pnpm-workspace.yaml | 5 +++++ 2 files changed, 13 insertions(+), 5 deletions(-) create mode 100644 pnpm-workspace.yaml diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml index fd73b0b..7918e7a 100644 --- a/pnpm-lock.yaml +++ b/pnpm-lock.yaml @@ -4,6 +4,9 @@ settings: autoInstallPeers: true excludeLinksFromLockfile: false +overrides: + brace-expansion: 5.0.9 + importers: .: @@ -189,9 +192,9 @@ packages: resolution: {integrity: sha512-BLrgEcRTwX2o6gGxGOCNyMvGSp35YofuYzw9h1IMTRmKqttAZZVU67bdb9Pr2vUHA8+j3i2tJfjO6C6+4myGTA==} engines: {node: 18 || 20 || >=22} - brace-expansion@5.0.5: - resolution: {integrity: sha512-VZznLgtwhn+Mact9tfiwx64fA9erHH/MCXEUfB/0bX/6Fz6ny5EGTXYltMocqg4xFAQZtnO3DHWWXi8RiuN7cQ==} - engines: {node: 18 || 20 || >=22} + brace-expansion@5.0.9: + resolution: {integrity: sha512-ScQ4IuvIEF1TMlP7Zt+vjJ//9zlPb2SDcxWxM3bk8s6t6GGdJ7KO1dCcTidOPJKePW30LE/2cT7wCyPho9/Wxg==} + engines: {node: 20 || >=22} cross-spawn@7.0.6: resolution: {integrity: sha512-uV2QOWP2nWzsy2aMp8aRibhi9dlzF5Hgh5SHaB9OiTGEyDTiJJyx0uy51QXdyWbtAHNua4XJzUKca3OzKUd3vA==} @@ -627,7 +630,7 @@ snapshots: balanced-match@4.0.4: {} - brace-expansion@5.0.5: + brace-expansion@5.0.9: dependencies: balanced-match: 4.0.4 @@ -779,7 +782,7 @@ snapshots: minimatch@10.2.4: dependencies: - brace-expansion: 5.0.5 + brace-expansion: 5.0.9 ms@2.1.3: {} diff --git a/pnpm-workspace.yaml b/pnpm-workspace.yaml new file mode 100644 index 0000000..dbfcdbf --- /dev/null +++ b/pnpm-workspace.yaml @@ -0,0 +1,5 @@ +packages: + - '.' + +overrides: + brace-expansion: 5.0.9 From 19998de42bdbd2bc54eea1c69794a1fb937b3791 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 19:56:49 +0200 Subject: [PATCH 08/66] chore: adopt NovaForge compute contract Verification: remote-compute-audit; compute guard heavy/light refusal checks; bash -n; git diff --check --- .vscode/tasks.json | 67 ++++++++++++ AGENTS.md | 72 +++++++++++++ README.md | 6 ++ docs/compute-distribution.md | 129 ++++++++++++++++++++++ docs/novaforge-ax102-stack.md | 84 +++++++++++++++ docs/operator-tooling.md | 150 ++++++++++++++++++++++++++ docs/remote-workspace.md | 63 +++++++++++ scripts/check-compute-distribution.sh | 123 +++++++++++++++++++++ 8 files changed, 694 insertions(+) create mode 100644 .vscode/tasks.json create mode 100644 AGENTS.md create mode 100644 docs/compute-distribution.md create mode 100644 docs/novaforge-ax102-stack.md create mode 100644 docs/operator-tooling.md create mode 100644 docs/remote-workspace.md create mode 100755 scripts/check-compute-distribution.sh diff --git a/.vscode/tasks.json b/.vscode/tasks.json new file mode 100644 index 0000000..2f86eb9 --- /dev/null +++ b/.vscode/tasks.json @@ -0,0 +1,67 @@ +{ + "version": "2.0.0", + "tasks": [ + { + "label": "Compute Guard: Check command", + "type": "shell", + "command": "${workspaceFolder}/scripts/check-compute-distribution.sh \"${input:guardCommand}\"", + "problemMatcher": [] + }, + { + "label": "Remote Compute: Dry run command", + "type": "shell", + "command": "${env:HOME}/.codex/bin/remote-compute-run --dry-run -- ${input:remoteCommand}", + "problemMatcher": [] + }, + { + "label": "Remote Compute: Run command on Hetzner", + "type": "shell", + "command": "${env:HOME}/.codex/bin/remote-compute-run -- ${input:remoteCommand}", + "problemMatcher": [] + }, + { + "label": "Remote Compute: Provision workspace", + "type": "shell", + "command": "${env:HOME}/.codex/bin/remote-compute-provision", + "problemMatcher": [] + }, + { + "label": "Remote Compute: Git status on Hetzner", + "type": "shell", + "command": "${env:HOME}/.codex/bin/remote-compute-run --no-update -- git status --short --branch", + "problemMatcher": [] + }, + { + "label": "Remote Compute: Doctor", + "type": "shell", + "command": "${env:HOME}/.codex/bin/remote-compute-doctor --quick ${workspaceFolder}", + "problemMatcher": [] + }, + { + "label": "Remote Compute: Inventory", + "type": "shell", + "command": "${env:HOME}/.codex/bin/remote-compute-inventory ${workspaceFolder}", + "problemMatcher": [] + }, + { + "label": "Local Sovereignty Audit", + "type": "shell", + "command": "${env:HOME}/.codex/bin/local-sovereignty-audit", + "problemMatcher": [] + } + ], + "inputs": [ + { + "id": "guardCommand", + "type": "promptString", + "description": "Workload or local command to classify", + "default": "pnpm test" + }, + { + "id": "remoteCommand", + "type": "promptString", + "description": "Command to run on Hetzner from this repo's remote workspace", + "default": "git status --short --branch" + } + ] +} diff --git a/AGENTS.md b/AGENTS.md new file mode 100644 index 0000000..5a10597 --- /dev/null +++ b/AGENTS.md @@ -0,0 +1,72 @@ +# Remote Compute Distribution + +This project follows the Mac Mini / Hetzner compute boundary. + +## Execution Identity + +- `Mac_Mini`: local control surface only. +- `Hetzner_AX102`: remote metabolic infrastructure. + +## Mac Mini Allowed + +- VS Code GUI +- editing +- code navigation +- git status, diff, branch, commit, push, pull, PR metadata +- local secrets and signing material +- lightweight smoke checks that finish quickly and do not create sustained memory, swap, thermal, Docker, indexing, OCR, embedding, build, or test load + +## Hetzner Required + +- dependency installs +- Docker and devcontainers +- full test suites +- CI reproduction +- heavy builds and release packaging +- long-running dev servers +- database services +- embeddings, vectorization, semantic indexing +- OCR and batch document processing +- coverage generation +- benchmark gates +- background inference + +## Required Workflow + +Open the project through VS Code Remote SSH when doing real implementation: + +```text +VS Code GUI: Mac Mini +Workspace: Hetzner +Terminal: Hetzner +Heavy proof: GitHub Actions or Hetzner +``` + +Local Mac checkouts are for navigation, small edits, and emergency fixes only. Do not present heavy local execution as completion proof. + +## Proof Integrity + +Remote proof must correspond to explicit project state. If the local worktree is +dirty, `remote-compute-run` refuses by default because the Hetzner workspace may +represent stale code. Commit or stash the local changes, use +`remote-compute-provision` for a clean bundle, or use `--allow-dirty-local` +only as an explicit human-approved non-proof override. + +## Versionability + +The compute-distribution contract surfaces must be versionable: + +- `docs/compute-distribution.md` +- `docs/operator-tooling.md` +- `docs/remote-workspace.md` +- `docs/novaforge-ax102-stack.md` +- `scripts/check-compute-distribution.sh` +- `AGENTS.md` +- `.vscode/tasks.json` + +Do not hide these files behind `.gitignore`. `remote-compute-audit` treats +ignored contract files as invalid even when they exist locally. + +## Override + +Local heavy compute requires explicit human override for the specific task. Hidden fallback to the Mac Mini is invalid. diff --git a/README.md b/README.md index f096d93..ced6f45 100644 --- a/README.md +++ b/README.md @@ -189,3 +189,9 @@ Current CI enforces: - typecheck (`pnpm typecheck`) - build (`pnpm build`) - baseline tests (`pnpm test`) + +## Compute Distribution + +This project follows the Mac Mini / Hetzner split documented in `docs/compute-distribution.md`. +Use VS Code Remote SSH for normal implementation and use GitHub Actions or Hetzner output as proof for heavy gates. +See `docs/operator-tooling.md` and `docs/remote-workspace.md` for the operator contract and remote workspace proof rules. diff --git a/docs/compute-distribution.md b/docs/compute-distribution.md new file mode 100644 index 0000000..3d0412d --- /dev/null +++ b/docs/compute-distribution.md @@ -0,0 +1,129 @@ +# Compute Distribution + +This project uses the global Mac Mini / Hetzner workload split. + +## Rule + +The Mac Mini is the local control surface. Hetzner is the compute substrate. + +## Local Safe + +- VS Code GUI +- code reading and editing +- `git status` +- `git diff` +- `git log` +- small targeted smoke checks +- secrets/signing operations that must remain local + +## Remote Required + +- dependency installation +- Docker and devcontainers +- full test suites +- CI reproduction +- release builds and packaging +- coverage +- benchmarks +- OCR +- embeddings +- vector indexing +- long-running servers +- databases +- background inference + +## Operator Workflow + +Use VS Code Remote SSH for normal work: + +```text +Remote host: hetzner-server +Remote workspace: /home/krille/workspaces/selectpilot +``` + +Use GitHub Actions or Hetzner terminal output as proof for heavy gates. Local Mac output is valid only for bounded smoke checks. + +Remote proof is valid only when it corresponds to the explicit project state. +If the local worktree is dirty, `remote-compute-run` must refuse by default +because the Hetzner workspace may represent stale code. Commit or stash the +local changes, use `remote-compute-provision` for a clean bundle, or make an +explicit human-approved non-proof override with `--allow-dirty-local`. + +Dirty protected repositories require read-only reconciliation evidence before +any merge, stash, branch, reset, or manual integration strategy is selected. +Use `remote-compute-reconcile-report` to surface upstream delta, dirty count, +local-only contract files, changed paths, and a recommended next step without +mutating the worktree. The same evidence is available with +`remote-compute-reconcile-report --json`. +Use `remote-compute-inventory-summary` before broad adoption decisions to expose +protected, dirty-protected, external, container/no-origin, and auto-adopt +candidate counts without mutating repositories. +Use `remote-compute-inventory-action-plan` to convert inventory state into a +read-only per-repository action plan before deciding whether a repo needs +reconcile approval, fork/ownership decision, project identity decision, or no +action. +Use `remote-compute-inventory-decision-template` to turn unresolved action-plan +rows into a read-only, human-fillable decision template. Unresolved decisions +must remain `selected_decision: null` and `mutation_allowed: false` until a +separate human review fills the template. +Use `remote-compute-reconcile-plan` to inspect the read-only phase plan before +any human-approved local mutation. +Use `remote-compute-reconcile-preservation-plan` to inspect the dirty head, +dirty count, suggested preservation branch, verified-export requirement, and +human-approval requirement before any dirty protected repository is mutated. +Use `remote-compute-reconcile-export --out ` to write patches and evidence +outside the source repository before a manual reconcile. +Exported reconcile evidence must include checksums for replayability and +integrity review. +Verify exported reconcile evidence with +`remote-compute-reconcile-verify-export ` before manual application. +Run `remote-compute-reconcile-mutation-preflight ` before any +human-approved mutation to prove the verified export still matches current +dirty repository state. +Generate `remote-compute-reconcile-approval-template ` before any dirty +protected repository is mutated; every repository must remain `approved: false` +until explicit human review changes that approval outside this read-only step. +Verify the human-filled approval file with +`remote-compute-reconcile-verify-approval ` before any +mutation; approval is invalid unless each repository row is explicitly approved +with reason, approver, and timestamp. +Generate `remote-compute-reconcile-mutation-packet ` after +approval verification and before mutation; the packet must remain read-only and +surface exact preservation branches, dirty heads, evidence paths, and +operator-visible commands. + +## Versionability + +These contract surfaces must be versionable and must not be hidden by +`.gitignore`: + +- `docs/compute-distribution.md` +- `docs/operator-tooling.md` +- `docs/remote-workspace.md` +- `docs/novaforge-ax102-stack.md` +- `scripts/check-compute-distribution.sh` +- `AGENTS.md` +- `.vscode/tasks.json` + +`remote-compute-audit` fails closed if any of these files exist only as ignored +local state. + +## NovaForge Stack + +This project docks into the AX102 stack documented in +`docs/novaforge-ax102-stack.md`. + +## Operator Tooling + +`docs/operator-tooling.md` defines the local tools that enforce this boundary. +`docs/remote-workspace.md` defines the Hetzner workspace and proof rules. + +## Local Guard + +Run: + +```bash +scripts/check-compute-distribution.sh "command or workload description" +``` + +The guard exits non-zero for known heavy local workloads. diff --git a/docs/novaforge-ax102-stack.md b/docs/novaforge-ax102-stack.md new file mode 100644 index 0000000..0580ae0 --- /dev/null +++ b/docs/novaforge-ax102-stack.md @@ -0,0 +1,84 @@ +# NovaForge AX102 Stack + +NovaForge is the operational expression of the compute-distribution boundary on +Hetzner AX102. + +It is not generic hosting and must not become a remote desktop replacement. Its +topology is: + +```text +persistent sovereign metabolic layer +``` + +## Role + +NovaForge carries sustained thermodynamic load for Namaka systems: + +- dependency installation +- Docker and devcontainer execution +- full tests, heavy builds, coverage, benchmarks, and release packaging +- persistent PostgreSQL, Redis, ChromaDB, queues, replay engines, and runtime services +- OCR, embeddings, vectorization, chunking, semantic indexing, and batch processing +- monitoring, backup, and long-running service stability + +The Mac Mini remains the sovereign control surface for UI, editing, git +metadata, local secrets, signing material, orchestration approval, lightweight +smoke checks, and local reasoning. + +## Base System + +The AX102 baseline is Ubuntu Server 24.04 LTS. + +Required service classes are: + +- Docker Compose +- remote devcontainers +- PostgreSQL +- Redis +- ChromaDB +- queue workers +- OCR workers +- embedding workers +- semantic indexing workers +- reverse proxy +- monitoring +- backup + +Implementation may use different package managers or deployment automation, but +the resulting system must expose these service classes as explicit, inspectable +infrastructure. Hidden installation state is invalid. + +## First Priority Migration + +Move these workloads first: + +1. NovaArchive OCR +2. NovaArchive embeddings +3. NovaArchive vectorization +4. NovaArchive chunking +5. NovaArchive semantic indexing +6. ChromaDB remote persistence +7. PostgreSQL remote persistence +8. Redis remote persistence +9. queue workers +10. replay engines + +These workloads are metabolic load, not control-surface activity. + +## Forbidden Topologies + +NovaForge must not become: + +- a remote desktop replacement +- generic hosting +- silent local heavy fallback + +If AX102 is unavailable, the system must surface the causal failure or route to +another explicit remote proof surface. It must not silently pull heavy execution +back to the Mac Mini. + +## Docking Boundary + +This document governs the broader AX102 stack that downstream systems can dock +into. Axiom Core may validate the topology, but it is not itself a long-running +service. diff --git a/docs/operator-tooling.md b/docs/operator-tooling.md new file mode 100644 index 0000000..2ad6407 --- /dev/null +++ b/docs/operator-tooling.md @@ -0,0 +1,150 @@ +# Operator Tooling Contract + +This repository uses local operator tools to enforce the Mac Mini / Hetzner +compute boundary. + +The tools may live outside this repository on the operator machine, but their +behavior is part of the compute-distribution contract. A repository is not +protected merely because files exist locally; tooling must prove the surfaces, +route heavy work away from the Mac Mini, and fail closed when proof would be +stale. + +## Required Tools + +- `remote-compute-adopt` +- `remote-compute-audit` +- `remote-compute-inventory` +- `remote-compute-inventory-summary` +- `remote-compute-inventory-action-plan` +- `remote-compute-inventory-decision-template` +- `remote-compute-provision` +- `remote-compute-run` +- `remote-compute-reconcile-report` +- `remote-compute-reconcile-plan` +- `remote-compute-reconcile-preservation-plan` +- `remote-compute-reconcile-export` +- `remote-compute-reconcile-verify-export` +- `remote-compute-reconcile-mutation-preflight` +- `remote-compute-reconcile-approval-template` +- `remote-compute-reconcile-verify-approval` +- `remote-compute-reconcile-mutation-packet` +- `remote-compute-doctor` +- `local-sovereignty-audit` + +## Required Behavior + +`remote-compute-adopt` installs the project contract surfaces and must not +silently overwrite dirty existing files. + +`remote-compute-audit` fails closed when a required contract surface is missing, +incomplete, non-executable where execution is required, or hidden by +`.gitignore`. + +`remote-compute-inventory` reports repository state without mutating it. Owned +dirty repositories with valid contracts must be classified as +`protected_dirty_local_only`, not as ready remote proof targets. + +`remote-compute-inventory-summary` summarizes inventory coverage without +mutating repositories. It must report protected, dirty-protected, external, +container/no-origin, and auto-adopt candidate counts in machine-readable form +so operators can see which repositories are safe to touch. + +`remote-compute-inventory-action-plan` converts inventory state into a +read-only per-repository action plan. It must classify each repository as +already protected, requiring verified reconcile approval, requiring fork or +ownership decision, requiring project identity decision, or requiring manual +classification. It must not mutate repositories. + +`remote-compute-inventory-decision-template` converts the action plan into a +read-only, human-fillable decision template. It must default every unresolved +decision to `selected_decision: null`, `mutation_allowed: false`, and must not +approve adoption, create branches, or write files. + +`remote-compute-provision` is the clean-bundle path for remote workspace +provisioning. It must refuse unclean local worktrees unless the workflow +explicitly preserves or packages the state being provisioned. + +When the local worktree is clean, the generated bundle is the authoritative +remote workspace state. `remote-compute-provision` must update the remote branch +and tracking ref from the bundle even if the Hetzner workspace is otherwise +clean. A clean but stale remote workspace is invalid proof. + +`remote-compute-run` executes commands on Hetzner, not locally. It must refuse +dirty local worktrees by default because the remote workspace may represent +stale code. `--allow-dirty-local` is allowed only as an explicit +human-approved non-proof override. + +`remote-compute-reconcile-report` reports dirty protected repositories before +any integration strategy is selected. It must be read-only and must surface the +branch, upstream delta, dirty count, local-only contract files, and changed +paths so the human can choose merge, stash, branch, or manual reconciliation +without losing local state. It must also expose a recommended next step and a +machine-readable JSON form so downstream operators can review the strategy +without parsing terminal prose. + +`remote-compute-reconcile-plan` converts the read-only report into an explicit +read-only sequence of strategy phases. It must not execute merge, stash, reset, +checkout, add, commit, or file writes; it only prepares the human-approved +reconcile path. + +`remote-compute-reconcile-preservation-plan` converts the read-only report into +explicit preservation preconditions before any dirty protected repository is +mutated. It must emit the dirty head, dirty count, preservation requirement, +suggested preservation branch, verified-export requirement, human-approval +requirement, and read-only flag without creating branches or writing files. + +`remote-compute-reconcile-export` writes reconcile evidence to an explicit +external artifact directory. It may export status, patches, upstream diffs, and +local contract-file copies, but it must not write inside source repositories or +execute merge, stash, reset, checkout, add, or commit. It must include +checksums for exported artifacts so the reconcile evidence can be verified +before manual application. + +`remote-compute-reconcile-verify-export` verifies an exported reconcile artifact +directory before any manual application. It must check required files, +manifest rows, JSON structure, read-only plan flags, and checksums. + +`remote-compute-reconcile-mutation-preflight` verifies that an exported +reconcile artifact still matches the current dirty repository state before any +human-approved mutation. It must fail closed when the dirty head, branch, dirty +count, preservation flags, or export checksums drift. + +`remote-compute-reconcile-approval-template` emits a read-only, machine-readable +approval template after export verification and mutation preflight pass. It must +default every repository to `approved: false` and must not create branches or +write files. + +`remote-compute-reconcile-verify-approval` verifies a human-filled approval file +against the exported reconcile artifact and current mutation preflight. It must +fail closed unless every repository row is approved with non-empty approval +reason, approver, and approval timestamp. It must not create branches or write +files. + +`remote-compute-reconcile-mutation-packet` emits the final read-only mutation +packet after approval verification succeeds. It may describe preservation branch +commands and evidence paths, but it must not create branches, write files, or +execute merge, stash, reset, checkout, add, or commit. + +`remote-compute-doctor` verifies local tooling, inventory, remote reachability, +and standard remote workspaces without turning a failure into local heavy +execution. + +`local-sovereignty-audit` verifies the full operator surface: + +- shell syntax of operator tools +- constitutional compute-distribution policy +- protected repository contracts +- heavy-workload guard refusal +- light git operation allowance +- VS Code remote-compute tasks +- remote workspace health +- dirty-worktree refusal in `remote-compute-run` + +## Proof Rule + +Heavy completion proof must come from GitHub Actions or Hetzner. Local operator +tool output is valid only for routing, contract inspection, bounded smoke +checks, and refusal proof. + +If a remote proof surface is unavailable, the correct result is a visible causal +failure. The tools must not silently execute heavy work on the Mac Mini. diff --git a/docs/remote-workspace.md b/docs/remote-workspace.md new file mode 100644 index 0000000..aad2d6b --- /dev/null +++ b/docs/remote-workspace.md @@ -0,0 +1,63 @@ +# Remote Workspace + +This repository's remote metabolic workspace is: + +```text +Host: hetzner-server +Workspace: /home/krille/workspaces/selectpilot +``` + +The Mac Mini may open the workspace through VS Code Remote SSH, but heavy +commands must execute on Hetzner or GitHub Actions. + +## Valid Heavy Proof + +Heavy proof is valid only when: + +- the remote workspace branch matches the explicit project state +- the remote workspace is clean +- the command output comes from Hetzner or GitHub Actions +- local dirty work has not been used as completion proof + +## Invalid Heavy Proof + +The following are invalid: + +- local Mac output for full tests, builds, benchmarks, OCR, vectorization, + indexing, Docker, devcontainers, or persistent services +- clean but stale Hetzner workspaces +- hidden fallback from remote execution to local execution +- proof generated from uncommitted local state unless explicitly packaged and + named as non-mainline proof + +## Standard Commands + +Inspect without mutation: + +```bash +remote-compute-run --dry-run -- git status --short --branch +``` + +Provision the remote workspace from a clean local checkout: + +```bash +remote-compute-provision +``` + +Run a remote command: + +```bash +remote-compute-run -- pnpm test +``` + +Audit the contract: + +```bash +remote-compute-audit . +``` + +Run the operator health check: + +```bash +remote-compute-doctor --quick . +``` diff --git a/scripts/check-compute-distribution.sh b/scripts/check-compute-distribution.sh new file mode 100755 index 0000000..9bdfcbb --- /dev/null +++ b/scripts/check-compute-distribution.sh @@ -0,0 +1,123 @@ +#!/bin/bash +set -euo pipefail + +description="${*:-}" + +if [[ -z "$description" ]]; then + cat >&2 <<'USAGE' +Usage: + scripts/check-compute-distribution.sh "command or workload description" + +The Mac Mini is a control surface. Heavy workloads must run on Hetzner or GitHub Actions. +USAGE + exit 64 +fi + +normalized="$(printf '%s' "$description" | tr '[:upper:]' '[:lower:]')" + +heavy_patterns=( + "pnpm install" + "pnpm i" + "pnpm test" + "pnpm -r test" + "pnpm build" + "pnpm validate" + "pnpm benchmark" + "pnpm harness" + "pnpm render" + "render:review" + "npm install" + "npm ci" + "npm test" + "npm run test" + "npm run build" + "npm run validate" + "corepack enable" + "yarn install" + "yarn test" + "yarn build" + "bun install" + "bun test" + "bun run build" + "uv sync" + "uv pip install" + "poetry install" + "swift build" + "swift test" + "xcodebuild" + "cargo build" + "cargo test" + "cargo install" + "go test ./..." + "python -m pip install" + "pip install" + "python -m compileall" + "pytest" + "unittest discover" + "pip-audit" + "security_audit" + "pre-commit run --all-files" + "docker build" + "docker compose" + "docker run" + "docker pull" + "docker system" + "docker image" + "docker volume" + "devcontainer" + "devcontainers" + "coverage" + "benchmark" + "ocr" + "tesseract" + "poppler" + "pdfplumber" + "pymupdf" + "embedding" + "sentence-transformers" + "transformers" + "torch" + "accelerate" + "vector" + "vectorization" + "indexing" + "chroma" + "chromadb" + "postgres" + "postgresql" + "redis" + "redis-server" + "celery" + "meilisearch" + "tantivy" + "release" + "package" + "backtest" + "simulation" + "prediction" + "replay" + "schema validation" + "full registry" + "organization-wide" + "generated registry" +) + +for pattern in "${heavy_patterns[@]}"; do + if [[ "$normalized" == *"$pattern"* ]]; then + cat >&2 < Date: Tue, 18 Aug 2026 20:57:28 +0200 Subject: [PATCH 09/66] fix(runtime): close profile acceptance gaps --- panel/runtime-profiles.js | 2 +- panel/runtime-profiles.ts | 2 +- server/ollama_client.py | 31 ++++++++++++++++----------- server/runtime_profiles.py | 2 +- tests/server/test_ollama_client.py | 20 +++++++++++++++++ tests/server/test_runtime_profiles.py | 27 +++++++++++++++++++++++ 6 files changed, 69 insertions(+), 15 deletions(-) diff --git a/panel/runtime-profiles.js b/panel/runtime-profiles.js index 8bba413..399b6ba 100644 --- a/panel/runtime-profiles.js +++ b/panel/runtime-profiles.js @@ -26,7 +26,7 @@ export const RUNTIME_PROFILES = [ key: 'advanced', label: 'Advanced', description: 'Manual opt-in profile for stronger reasoning on larger machines.', - generation_model: 'gemma4:12b-it-qat', + generation_model: 'qwen2.5:7b', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 32768, target_latency: '4-10s', diff --git a/panel/runtime-profiles.ts b/panel/runtime-profiles.ts index 228d721..65f8762 100644 --- a/panel/runtime-profiles.ts +++ b/panel/runtime-profiles.ts @@ -41,7 +41,7 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ key: 'advanced', label: 'Advanced', description: 'Manual opt-in profile for stronger reasoning on larger machines.', - generation_model: 'gemma4:12b-it-qat', + generation_model: 'qwen2.5:7b', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 32768, target_latency: '4-10s', diff --git a/server/ollama_client.py b/server/ollama_client.py index 60e62f4..3537bfa 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -101,32 +101,39 @@ class OllamaError(RuntimeError): pass +def _positive_int(value: Any, name: str) -> int: + try: + parsed = int(value) + except (TypeError, ValueError) as exc: + raise OllamaError(f"{name} must be a positive integer") from exc + if parsed <= 0: + raise OllamaError(f"{name} must be a positive integer") + return parsed + + class OllamaClient: def __init__(self, config: OllamaConfig | None = None): if config is None: - default_generation_model = "llama3.2" + default_generation_model = "gemma4:e2b-it-qat" default_embed_model = "nomic-embed-text-v2-moe:latest" default_num_ctx = 16384 runtime_profile = os.environ.get("CHROMEAI_RUNTIME_PROFILE", "auto") - try: - from runtime_profiles import get_runtime_profile, recommend_runtime_profile + from runtime_profiles import get_runtime_profile, recommend_runtime_profile - recommendation = recommend_runtime_profile() - resolved_profile = recommendation["recommended_profile"] if runtime_profile == "auto" else runtime_profile - profile = get_runtime_profile(resolved_profile) - default_generation_model = profile.generation_model - default_embed_model = profile.embedding_model - default_num_ctx = profile.num_ctx - except Exception: - pass + recommendation = recommend_runtime_profile() + resolved_profile = recommendation["recommended_profile"] if runtime_profile == "auto" else runtime_profile + profile = get_runtime_profile(resolved_profile) + default_generation_model = profile.generation_model + default_embed_model = profile.embedding_model + default_num_ctx = profile.num_ctx config = OllamaConfig( base_url=_normalize_base_url(os.environ.get("CHROMEAI_OLLAMA_BASE_URL", "http://127.0.0.1:11434")), model=os.environ.get("CHROMEAI_OLLAMA_MODEL", default_generation_model), embed_model=os.environ.get("CHROMEAI_OLLAMA_EMBED_MODEL", default_embed_model), timeout_seconds=float(os.environ.get("CHROMEAI_OLLAMA_TIMEOUT_SECONDS", "30")), - num_ctx=int(os.environ.get("CHROMEAI_OLLAMA_NUM_CTX", str(default_num_ctx))), + num_ctx=_positive_int(os.environ.get("CHROMEAI_OLLAMA_NUM_CTX", default_num_ctx), "CHROMEAI_OLLAMA_NUM_CTX"), ) self.config = config diff --git a/server/runtime_profiles.py b/server/runtime_profiles.py index 97496cf..e1ad11b 100644 --- a/server/runtime_profiles.py +++ b/server/runtime_profiles.py @@ -47,7 +47,7 @@ class RuntimeProfile: key="advanced", label="Advanced", description="Manual opt-in profile for stronger reasoning on larger machines.", - generation_model="gemma4:12b-it-qat", + generation_model="qwen2.5:7b", embedding_model="nomic-embed-text-v2-moe:latest", num_ctx=32768, target_latency="4-10s", diff --git a/tests/server/test_ollama_client.py b/tests/server/test_ollama_client.py index b49c990..478daab 100644 --- a/tests/server/test_ollama_client.py +++ b/tests/server/test_ollama_client.py @@ -58,6 +58,26 @@ def test_summarize_rejects_missing_required_fields(self) -> None: with self.assertRaisesRegex(OllamaError, "missing required fields"): client.summarize("content") + def test_all_structured_routes_send_closed_object_schemas(self) -> None: + responses = [ + {"response": '{"summary":"S","bullets":[],"action_items":[],"title":"T","tags":[]}'}, + {"response": '{"reasoning":[],"markdown":"M","json":{}}'}, + {"response": '{"summary":"S","action_items":[],"decisions":[],"risks":[],"follow_ups":[]}'}, + ] + client = StubOllamaClient(responses.pop(0)) + payloads = [] + for operation in ( + lambda: client.summarize("content"), + lambda: client.agent("prompt"), + lambda: client.extract("content", preset_key="action_brief"), + ): + operation() + payloads.append(client.last_payload) + if responses: + client.response = responses.pop(0) + self.assertEqual([payload["format"]["type"] for payload in payloads], ["object"] * 3) + self.assertTrue(all(payload["format"]["additionalProperties"] is False for payload in payloads)) + if __name__ == "__main__": unittest.main() diff --git a/tests/server/test_runtime_profiles.py b/tests/server/test_runtime_profiles.py index 59eb57a..9a7418a 100644 --- a/tests/server/test_runtime_profiles.py +++ b/tests/server/test_runtime_profiles.py @@ -2,6 +2,7 @@ import sys import unittest +from unittest.mock import patch from pathlib import Path ROOT = Path(__file__).resolve().parents[2] @@ -10,6 +11,7 @@ sys.path.insert(0, str(SERVER_DIR)) from runtime_profiles import build_bootstrap_commands, get_runtime_profile, recommend_runtime_profile # noqa: E402 +from ollama_client import OllamaClient, OllamaError # noqa: E402 class RuntimeProfileTests(unittest.TestCase): @@ -25,6 +27,31 @@ def test_auto_recommendation_prefers_balanced_on_large_machines(self) -> None: ) self.assertEqual(recommendation["recommended_profile"], "balanced") + def test_low_memory_and_intel_hardware_use_fast(self) -> None: + low_memory = recommend_runtime_profile( + {"machine": "arm64", "memory_gb": 8, "platform": "darwin", "cpu_count": 8} + ) + intel = recommend_runtime_profile( + {"machine": "x86_64", "memory_gb": 64, "platform": "darwin", "cpu_count": 12} + ) + self.assertEqual(low_memory["recommended_profile"], "fast") + self.assertEqual(intel["recommended_profile"], "fast") + + def test_explicit_profile_and_context_override_are_authoritative(self) -> None: + with patch.dict("os.environ", { + "CHROMEAI_RUNTIME_PROFILE": "balanced", + "CHROMEAI_OLLAMA_MODEL": "custom-local:model", + "CHROMEAI_OLLAMA_NUM_CTX": "24576", + }, clear=True): + client = OllamaClient() + self.assertEqual(client.config.model, "custom-local:model") + self.assertEqual(client.config.num_ctx, 24576) + + def test_invalid_context_override_fails_explicitly(self) -> None: + with patch.dict("os.environ", {"CHROMEAI_OLLAMA_NUM_CTX": "0"}, clear=True): + with self.assertRaisesRegex(OllamaError, "must be a positive integer"): + OllamaClient() + def test_bootstrap_command_contains_profile(self) -> None: command = build_bootstrap_commands("balanced", ROOT) self.assertIn("--profile balanced", command["command"]) From 32b55cd18c7ce83c728f0a1213a150582b2128ba Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 20:59:34 +0200 Subject: [PATCH 10/66] docs(runtime): record model evidence boundaries --- docs/model-runtime-decisions.md | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) create mode 100644 docs/model-runtime-decisions.md diff --git a/docs/model-runtime-decisions.md b/docs/model-runtime-decisions.md new file mode 100644 index 0000000..e283e31 --- /dev/null +++ b/docs/model-runtime-decisions.md @@ -0,0 +1,21 @@ +# Model runtime decisions + +## Current production defaults + +- Fast / unknown / under 16 GB / Intel: `gemma4:e2b-it-qat`, `num_ctx=16384`. +- Balanced / Apple Silicon with 16 GB or more: `gemma4:e4b-it-qat`, `num_ctx=32768`. +- Advanced: manual opt-in only, `qwen2.5:7b`, `num_ctx=32768`. +- Embeddings: `nomic-embed-text-v2-moe:latest`. + +Explicit environment overrides remain authoritative and must fail visibly when invalid or unavailable. SelectPilot never substitutes a different installed model, downloads a model without consent, or routes generation to a cloud model. + +## Deferred decisions + +- Keep `nomic-embed-text-v2-moe:latest` until measured retrieval quality makes embeddings a product bottleneck (SOD-397/SOD-401). +- Do not promote a larger Gemma 4 model for 32 GB machines until real M1 Max evidence qualifies its latency, memory use, structured-output validity, and failure behavior (SOD-398). +- Do not adopt Qwen3 4B for long context without actual usage evidence that 32K is insufficient (SOD-399). +- Do not replace Ollama with MLX/LM Studio without target-hardware measurements showing a material full-system advantage (SOD-400). + +## Evidence boundary + +Hetzner and GitHub Actions verify deterministic contracts, build integrity, privacy boundaries, and browser/server E2E behavior. They do not qualify Apple Silicon latency or memory claims. Target-hardware promotion remains blocked until evidence is captured on the named machine class under the repository benchmark contract. From e98fe1cb5b359518b78a2fc1df68cdc721e03803 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 21:06:29 +0200 Subject: [PATCH 11/66] fix(extension): restore real selected-text flow --- README.md | 18 +- ROADMAP.md | 52 ++-- content/content-script.bundle.js | 98 +++++++ manifest.json | 2 +- package.json | 3 +- panel/panel.js | 4 +- panel/panel.ts | 4 +- pnpm-lock.yaml | 271 ++++++++++++++++++++ pnpm-workspace.yaml | 3 + scripts/bootstrap-macos-local.sh | 38 ++- scripts/build-extension.mjs | 19 ++ server/runtime_profiles.py | 6 +- tests/e2e/extension-first-run-flow.spec.mjs | 72 +++++- tests/server/test_runtime_profiles.py | 7 +- 14 files changed, 537 insertions(+), 60 deletions(-) create mode 100644 content/content-script.bundle.js create mode 100644 scripts/build-extension.mjs diff --git a/README.md b/README.md index ced6f45..dc5ed3f 100644 --- a/README.md +++ b/README.md @@ -40,16 +40,16 @@ SelectPilot is built to keep the core workflow local-first and inspectable. ## Quick Start -1. Install Ollama -2. Bootstrap local runtime + extension build: +1. From the repository folder, run the one-command setup: ```bash pnpm setup:local -pnpm build ``` -3. Load unpacked extension in `chrome://extensions` -4. Select text → open side panel → click **Extract JSON** +This installs dependencies, builds the extension, installs/starts Ollama when needed, selects a hardware-safe local profile, pulls the exact models, installs the LaunchAgent, and refuses to finish unless the bridge and both configured models are healthy. + +2. Open `chrome://extensions`, enable **Developer mode**, choose **Load unpacked**, and select this repository folder. +3. Highlight text on a page → open SelectPilot → click **Extract JSON**. Optional local checks: @@ -58,6 +58,14 @@ curl http://127.0.0.1:8083/health pnpm test:privacy ``` +If setup stops, no fallback was applied. Copy/paste these diagnostics: + +```bash +tail -n 80 ~/Library/Logs/SelectPilot/nano.err +tail -n 80 ~/Library/Logs/SelectPilot/nano.log +curl -sSf http://127.0.0.1:8083/health +``` + --- ## Architecture diff --git a/ROADMAP.md b/ROADMAP.md index ef2394e..5825a2d 100644 --- a/ROADMAP.md +++ b/ROADMAP.md @@ -32,8 +32,8 @@ ### 4) Onboarding-friktion är för hög -- **Status:** 🟡 Delvis löst (bootstrap-script finns). -- **Åtgärd kvar:** ge ett verkligt one-command bootstrap-flöde som fungerar utan manuell felsökning. +- **Status:** ✅ Löst för det lokala macOS-flödet. +- **Implementerat:** `pnpm setup:local` bygger extensionen, installerar/startar Ollama vid behov, väljer hårdvaruprofil, hämtar exakta modeller, installerar LaunchAgent och failar om exakt bridge-/modellhälsa inte kan verifieras. ### 5) Nginx/hosts-topologi bör vara osynlig @@ -42,15 +42,15 @@ ### 6) README saknar tydligt 5-minutersflöde högst upp -- **Status:** 🟡 Delvis löst (sektionen "Local setup" är bra men inte tydligt märkt som snabbstart). -- **Åtgärd kvar:** lägg en kort **Quick Start (5 min)** tidigt i README med exakta steg i ordning. +- **Status:** ✅ Löst. +- **Implementerat:** trestegs Quick Start tidigt i README, ett enda setup-kommando och copy/paste-diagnostik utan dold fallback. ### 7) Inga end-to-end-tester för extension-flödet -- **Status:** 🟡 Delvis löst. +- **Status:** ✅ Löst för det primära selected-text-flödet. - **Implementerat:** Playwright-konfiguration och E2E-suite för runtime/privacy + mockad selected-text response shape. - **Implementerat nu även:** harness-baserat sidepanel E2E-flöde som verifierar privacy-indikator + lokal fetch-trafik i panelens användarflöde. -- **Åtgärd kvar:** utöka till full browser-extension user flow (markera text → sidepanel action → renderat svar) utan harness. +- **Implementerat:** riktig unpacked-extension-path med injicerat content script: markera text på webbsida → background/context → sidepanel-action → renderat svar, utan harness. --- @@ -58,12 +58,12 @@ ### 8) Inga automatiska regressionstester för privacy boundary -- **Status:** 🟡 Delvis löst. +- **Status:** ✅ Löst för kärnflödet. - **Implementerat:** - servertest för privacy proof (`tests/server/test_privacy_proof.py`) - Playwright-test som verifierar local-only privacy proof och local endpoints - Playwright-test (`tests/e2e/panel-no-leakage.spec.mjs`) som assertar lokal fetch-trafik + subtil visuell privacy proof i panel-harness -- **Åtgärd kvar:** utöka med strikt nätverks-assert på full extensionnivå (ingen extern trafik i hela sidepanel-flödet utan harness). +- **Implementerat:** full-extension-testet registrerar all HTTP(S)-trafik under kärnoperationen och kräver noll externa anrop. ### 9) Privacy-påstående är inte synligt verifierbart för ny användare @@ -107,13 +107,8 @@ ### Nästa sprint (måste först) -1. **P1.7** Full extension E2E (markera text → sidepanel action → svar) -2. **P1.8** Strikt nätverksregressionstest för hela extension-flödet -3. **P1.6** Tydlig Quick Start (5 min) i README-toppen - -### Därefter (releasekvalitet) - -4. **P1.4** Minska onboarding-friktion ytterligare i bootstrap +1. Verifiera one-command onboarding och verklig Ollama-inferens på namngiven Apple Silicon-målmaskin. +2. Samla långinput-, minnes- och latensevidens för Fast/Balanced innan slutlig modellpromotion. ### Strategisk polish @@ -131,18 +126,12 @@ ### 1) One-command installation (absolut först) - **Mål:** användaren ska klara onboarding med en enda kommando-rad. -- **Kvar att leverera:** - - gör `pnpm bootstrap:local` helt självbärande med robust felhantering - - automatisk kontroll av Ollama, modeller, launchd och `/health` - - tydlig slutrapport med exakt "klart/återstår" och nästa steg +- **Status:** Implementerat och kontraktstestat; verklig målmaskinsverifiering återstår. ### 2) "89-åring-läge" i README (ultrakort onboarding) - **Mål:** inga tekniska beslut i första flödet. -- **Kvar att leverera:** - - 3-stegs quick start högst upp i README (kopiera kommando → ladda extension → klart) - - separera nybörjarflöde från advanced/dev-sektioner - - lägg till "om något går fel" med copy/paste-kommandon +- **Status:** Trestegsflöde och copy/paste-diagnostik implementerade; användbarhetsvalidering på ren målmaskin återstår. ### 3) UI/UX-polish till premium-kvalitet @@ -155,10 +144,7 @@ ### 4) E2E-bevis för verkligt användarflöde - **Mål:** kunna verifiera funktion och privacy med reproducerbara tester. -- **Kvar att leverera:** - - deterministisk start/stop av lokal server i testflödet - - full extension-path: markera text → sidepanel action → renderat svar - - strikt nätverksregression: ingen extern trafik i kärnflödet +- **Status:** ✅ Implementerat: deterministisk serverlivscykel, full extension-path och strikt extern nätverksassert. ### 5) Strategisk finish (efter kvalitet + onboarding) @@ -176,19 +162,19 @@ | 1 | Versionsdrift | P0 | Löst | Låg | | 2 | Dubbla lock-filer | P0 | Löst | Låg | | 3 | Ogiltiga manifest-fält | P0 | Löst | Låg | -| 4 | Onboarding-friktion | P1 | Delvis | Hög | +| 4 | Onboarding-friktion | P1 | Löst* | Hög | | 5 | Nginx/hosts osynlighet | P1 | Löst* | Hög | -| 6 | README Quick Start | P1 | Delvis | Låg | -| 7 | E2E-tester extension | P1 | Delvis | Medium | -| 8 | Privacy regressionstester | P1 | Delvis | Medium | +| 6 | README Quick Start | P1 | Löst | Låg | +| 7 | E2E-tester extension | P1 | Löst | Medium | +| 8 | Privacy regressionstester | P1 | Löst | Medium | | 9 | Verifierbar privacy-yta | P1 | Löst | Medium | | 10 | Omformulera positioning | P2 | Delvis | Låg | | 11 | Utbyggbara presets | P2 | Delvis | Medium | | 12 | Team/self-hosted-path | P2 | Delvis | Låg | | 13 | Tier/pricing-integritet | P2 | Delvis | Medium | -\* Löst i huvudflödet, men behöver fortsatt skyddas från regression i docs/scripts. +\* Kod- och kontraktsnivå är löst; verklig ren målmaskinsverifiering återstår före fullsystems-DoD. --- -*Senast uppdaterad: 2026-03-25* \ No newline at end of file +*Senast uppdaterad: 2026-03-25* diff --git a/content/content-script.bundle.js b/content/content-script.bundle.js new file mode 100644 index 0000000..2577e4c --- /dev/null +++ b/content/content-script.bundle.js @@ -0,0 +1,98 @@ +"use strict"; +(() => { + // content/extract-text.js + function clean(text) { + return text.replace(/\s+/g, " ").trim(); + } + function extractSelection() { + const selection = window.getSelection(); + const text = selection ? selection.toString() : ""; + return { + text: clean(text), + url: location.href, + title: document.title + }; + } + function extractDocumentText() { + const article = document.querySelector("article"); + const target = article || document.body; + const text = clean(target.innerText || ""); + return { + text, + url: location.href, + title: document.title + }; + } + + // content/extract-audio.js + function extractAudio() { + const audio = document.querySelector("audio"); + if (!audio) + return null; + return { + audioUrl: audio.currentSrc || audio.src || void 0, + duration: Number.isFinite(audio.duration) ? audio.duration : void 0, + title: document.title, + pageUrl: location.href + }; + } + + // content/extract-video.js + async function captureFrame(video) { + try { + const canvas = document.createElement("canvas"); + canvas.width = video.videoWidth; + canvas.height = video.videoHeight; + const ctx = canvas.getContext("2d"); + if (!ctx) + return void 0; + ctx.drawImage(video, 0, 0, canvas.width, canvas.height); + return canvas.toDataURL("image/png"); + } catch (e) { + console.error("frame capture failed", e); + return void 0; + } + } + async function extractVideo() { + const video = document.querySelector("video"); + if (!video) + return null; + return { + poster: video.poster || void 0, + currentTime: video.currentTime, + duration: Number.isFinite(video.duration) ? video.duration : void 0, + frame: await captureFrame(video), + pageUrl: location.href + }; + } + + // utils/logger.js + var ENABLE_LOG = true; + function log(scope, ...args) { + if (!ENABLE_LOG) + return; + console.log(`[ChromeAI:${scope}]`, ...args); + } + + // content/content-script.ts + chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { + (async () => { + const response = {}; + if (msg.type === "content:get_selection") { + response.text = extractSelection(); + } + if (msg.type === "content:get_document") { + response.documentText = extractDocumentText(); + } + if (msg.type === "content:get_audio") { + response.audio = extractAudio(); + } + if (msg.type === "content:get_video") { + response.video = await extractVideo(); + } + sendResponse(response); + })(); + return true; + }); + log("content", "content script injected"); +})(); diff --git a/manifest.json b/manifest.json index 151521c..d26651d 100644 --- a/manifest.json +++ b/manifest.json @@ -33,7 +33,7 @@ "content_scripts": [ { "matches": [""], - "js": ["content/content-script.js"], + "js": ["content/content-script.bundle.js"], "run_at": "document_idle" } ], diff --git a/package.json b/package.json index 4df60b2..726db59 100644 --- a/package.json +++ b/package.json @@ -24,7 +24,7 @@ "runtime:rollback": "node ./scripts/runtime-rollback.mjs", "runtime:show-policy": "node ./scripts/runtime-show-policy.mjs", "sync:manifest-version": "node ./scripts/sync-manifest-version.mjs", - "build": "pnpm sync:manifest-version && tsc -p tsconfig.json", + "build": "pnpm sync:manifest-version && tsc -p tsconfig.json && node ./scripts/build-extension.mjs", "lint": "eslint --max-warnings=0 .", "lint:manifest": "node ./scripts/lint-manifest.mjs", "test:e2e": "playwright test --config tests/e2e/playwright.config.mjs", @@ -37,6 +37,7 @@ "@eslint/js": "^10.0.1", "@playwright/test": "^1.55.0", "@types/chrome": "^0.1.29", + "esbuild": "^0.28.2", "eslint": "^10.1.0", "globals": "^17.4.0", "typescript": "^5.9.2", diff --git a/panel/panel.js b/panel/panel.js index c87ab9c..40fbfac 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -115,9 +115,9 @@ const runtimeMetaOverlayState = { let intentSuggestions = []; let selectedIntentSuggestion = null; const FAST_INSTALL_COMMANDS = [ - 'ollama pull qwen2.5:0.5b', + 'ollama pull gemma4:e2b-it-qat', 'ollama pull nomic-embed-text-v2-moe:latest', - 'CHROMEAI_OLLAMA_MODEL=qwen2.5:0.5b ./scripts/install-macos-local.sh', + 'CHROMEAI_OLLAMA_MODEL=gemma4:e2b-it-qat CHROMEAI_OLLAMA_NUM_CTX=16384 ./scripts/install-macos-local.sh', ].join('\n'); const QUICK_SETUP_COMMANDS = [ 'pnpm setup:local', diff --git a/panel/panel.ts b/panel/panel.ts index 50cec24..be0b875 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -210,9 +210,9 @@ let intentSuggestions: string[] = []; let selectedIntentSuggestion: string | null = null; const FAST_INSTALL_COMMANDS = [ - 'ollama pull qwen2.5:0.5b', + 'ollama pull gemma4:e2b-it-qat', 'ollama pull nomic-embed-text-v2-moe:latest', - 'CHROMEAI_OLLAMA_MODEL=qwen2.5:0.5b ./scripts/install-macos-local.sh', + 'CHROMEAI_OLLAMA_MODEL=gemma4:e2b-it-qat CHROMEAI_OLLAMA_NUM_CTX=16384 ./scripts/install-macos-local.sh', ].join('\n'); const QUICK_SETUP_COMMANDS = [ diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml index 7918e7a..093c74b 100644 --- a/pnpm-lock.yaml +++ b/pnpm-lock.yaml @@ -20,6 +20,9 @@ importers: '@types/chrome': specifier: ^0.1.29 version: 0.1.38 + esbuild: + specifier: ^0.28.2 + version: 0.28.2 eslint: specifier: ^10.1.0 version: 10.1.0 @@ -35,6 +38,162 @@ importers: packages: + '@esbuild/aix-ppc64@0.28.2': + resolution: {integrity: sha512-XExcO+dvLKvVtNTibSTBej1NCAbaGhWn9Ww1ZPx80qsahhPFe/8jgWP0IchNe0F3HwkU7n8ejhH8bjonqht8mQ==} + engines: {node: '>=18'} + cpu: [ppc64] + os: [aix] + + '@esbuild/android-arm64@0.28.2': + resolution: {integrity: sha512-5YfKeeI8qWfBZIX+u2xZC3Zlb3Os/gLS2sbEKM+I4ZOcsWmHS2WLysCcQZDAFRslDUU5Oiq44gf6PYN1vGwG5A==} + engines: {node: '>=18'} + cpu: [arm64] + os: [android] + + '@esbuild/android-arm@0.28.2': + resolution: {integrity: sha512-kXXoiPVVGQcnIYGOeaovwOURpniDBpSq4A03qkQ+BMQqtGG6HYap3xne9C1O1yo4TR3qxlCX5IqqmX6fFo2Lqg==} + engines: {node: '>=18'} + cpu: [arm] + os: [android] + + '@esbuild/android-x64@0.28.2': + resolution: {integrity: sha512-O387ite7SzUyCcy3JQX4P4bLtEA7bLLkx+esve5JHnyYfNTxcVpXZo9jhdB0lTKN44gztELTdU7nS8Nr16Fs1Q==} + engines: {node: '>=18'} + cpu: [x64] + os: [android] + + '@esbuild/darwin-arm64@0.28.2': + resolution: {integrity: sha512-n4KqkOQrraxHJcgjM1RvwbigfQKIKJVpM7xp+KsxiyUSrRdIXnt73VhrPAx0fV44hgfmIVKjxMN9J1t5jySVkw==} + engines: {node: '>=18'} + cpu: [arm64] + os: [darwin] + + '@esbuild/darwin-x64@0.28.2': + resolution: {integrity: sha512-uq6suIWYP37qzGddBKPw5QEQPi6HiLGsO7UmkpfyaYNQ3D+rN6w6WfwH+nuqcGXWvawGwxOEroO4YGnFh95azw==} + engines: {node: '>=18'} + cpu: [x64] + os: [darwin] + + '@esbuild/freebsd-arm64@0.28.2': + resolution: {integrity: sha512-n+I0BTSRIoy+d6RPKnEVwql5UwBJolytvY4mAOIEJorKlqgPII8ix6slVVrfZ5Tnj7glIZvloylbB/EJPMWEXw==} + engines: {node: '>=18'} + cpu: [arm64] + os: [freebsd] + + '@esbuild/freebsd-x64@0.28.2': + resolution: {integrity: sha512-78XJTJkvPs0kz2w61301PJjXl4g7q3JqiYMZ/M/yVI73EHBrCRTgkhu9oqG7vPqq+a/yadEW8aD+agKlk5xrmg==} + engines: {node: '>=18'} + cpu: [x64] + os: [freebsd] + + '@esbuild/linux-arm64@0.28.2': + resolution: {integrity: sha512-pW4AC0P3it8c7do9MVM4p51FzHzdM/TZrerurgRcHJ2WTa1VQ1CIq18xncfpBJw4ojkiZZrKW2yIBWBP92j6Ug==} + engines: {node: '>=18'} + cpu: [arm64] + os: [linux] + + '@esbuild/linux-arm@0.28.2': + resolution: {integrity: sha512-XlDnu2q5yoqems+xay6wSAcg9DDD7K9RLKZEBOMZm3ckNpJBvOX20tSfby8KfrrhINDyv9V2YVZKY/SpoGJI8w==} + engines: {node: '>=18'} + cpu: [arm] + os: [linux] + + '@esbuild/linux-ia32@0.28.2': + resolution: {integrity: sha512-CYbnj78HsIeA+DhgUKgFCfvNsTHFhMMrinUrMZpDXJXKN8T3XViTZ/+wtHeVxEWY8ewSzTFN+nRmSwO2tZaLUQ==} + engines: {node: '>=18'} + cpu: [ia32] + os: [linux] + + '@esbuild/linux-loong64@0.28.2': + resolution: {integrity: sha512-buwkd8nsph4R+ajRvw0qM5Hja/TXQow3ptzWO2EbG/cqcIkHloRrdlBtQlshyYGTNFvfkfJ5tpPLVkY4DtsPfQ==} + engines: {node: '>=18'} + cpu: [loong64] + os: [linux] + + '@esbuild/linux-mips64el@0.28.2': + resolution: {integrity: sha512-ZVykbDyk7519VwiNb9Lcj9m8XM6v5V9uKPvrEMkkEedVewf+0itkhahp4HDpgERXhwLRpWFypsGbG/J8s0QjJA==} + engines: {node: '>=18'} + cpu: [mips64el] + os: [linux] + + '@esbuild/linux-ppc64@0.28.2': + resolution: {integrity: sha512-CAXl+Dtd9UUuJd8pKKdwh6MLm3MUMiqMPmhZ3tTSXPqfyQ3vDl6R5hZdZ/kYojK4ofXtdfSv1tFq8XzWx3heNQ==} + engines: {node: '>=18'} + cpu: [ppc64] + os: [linux] + + '@esbuild/linux-riscv64@0.28.2': + resolution: {integrity: sha512-GeXCej4IQtU1B+QlDV8W/RRvbzI3O/Stss+/bCXv4lZls5WGRtu2a+3JkA3i4qIUlMXpcHebWpF8AkJhATowuA==} + engines: {node: '>=18'} + cpu: [riscv64] + os: [linux] + + '@esbuild/linux-s390x@0.28.2': + resolution: {integrity: sha512-3H1weTYZPxt/WOhByszQZybS9w5lKzUn1FDMsgEChbHWQwHYQQRfBxgCcZvPhjHfKyJjIievvMmEUawJrdY9Dg==} + engines: {node: '>=18'} + cpu: [s390x] + os: [linux] + + '@esbuild/linux-x64@0.28.2': + resolution: {integrity: sha512-4xTZr1FUmSoQW4XIWmit3tzQrUTZM+N3P0XV8xROKYF50XfI7xeO90+1bZvNwxIufQ9hDQVRJH5YhgPVF8A/HQ==} + engines: {node: '>=18'} + cpu: [x64] + os: [linux] + + '@esbuild/netbsd-arm64@0.28.2': + resolution: {integrity: sha512-sSATRjPeDBg3pdgHoQfoYBob11Kk1FGa9lui5RIHZCoCkJa9QKlvl3/vKz2usCmYYjs7ymJR/2Nnsqe+Hjt5nw==} + engines: {node: '>=18'} + cpu: [arm64] + os: [netbsd] + + '@esbuild/netbsd-x64@0.28.2': + resolution: {integrity: sha512-lqnzCV+mM0gIADaKihiCg6ifgfU2L3h5E33rNQBN1Y4MaVGnzryzmvvf7UHxprpQdE8hpqLolJ9Rl+SkIRDpyw==} + engines: {node: '>=18'} + cpu: [x64] + os: [netbsd] + + '@esbuild/openbsd-arm64@0.28.2': + resolution: {integrity: sha512-AL2qJILH7lNjrDmCQDvdxMfAUIv8KMNZOvrwAQ8i8//ntL9FflhOyMJ8OZSMBb8/AWXe3/5v5S20y3zCoZWKoQ==} + engines: {node: '>=18'} + cpu: [arm64] + os: [openbsd] + + '@esbuild/openbsd-x64@0.28.2': + resolution: {integrity: sha512-QtiuPytchRyC4rwUKhexJdQKvDuZ6hWloi3igqPQNUJCS1/v9EiO3UTOXR6A3FoMo4fnAKbWJdqaIwhOzh8qEw==} + engines: {node: '>=18'} + cpu: [x64] + os: [openbsd] + + '@esbuild/openharmony-arm64@0.28.2': + resolution: {integrity: sha512-WkhYDmpTjLvGlScA1rwjRUmhl4k8oXR3cIbtqWmELgU/dFeHHlEllxDvdWcNJV9rbzCexB5vz8gtNewWLgCT7Q==} + engines: {node: '>=18'} + cpu: [arm64] + os: [openharmony] + + '@esbuild/sunos-x64@0.28.2': + resolution: {integrity: sha512-GPMSkTOtMnv2U2F8gxe4Io6qmVs+YKyp832Etqqxr0hFngmXQ3rzwytelm3GIn7T4VviRUlf3sOgBOiTdvaf7g==} + engines: {node: '>=18'} + cpu: [x64] + os: [sunos] + + '@esbuild/win32-arm64@0.28.2': + resolution: {integrity: sha512-PIhhEkE9uPBleRBrQEJpUn7MBnibZzbGzYWPmY3x+YoVg/95zbjB4CxPPOQ8l5tYYM4mMaCthF8/1DIfBQQyWQ==} + engines: {node: '>=18'} + cpu: [arm64] + os: [win32] + + '@esbuild/win32-ia32@0.28.2': + resolution: {integrity: sha512-YmJbfTlvU7Sdn9BB+4PRES4oB6pxgS37MAONj+hBr/cpXS1aBPKXxNnDbu+QCWPj0o9dgyxeq79g6c5P8KeuYA==} + engines: {node: '>=18'} + cpu: [ia32] + os: [win32] + + '@esbuild/win32-x64@0.28.2': + resolution: {integrity: sha512-5ebpxr3nWMzrL/rnUI755Jkuee0bHL/Gq0WTF9lvcpv73wAp5eu8MfBUgWK9bhWvZjj7yX8etf/8tI8Ney695g==} + engines: {node: '>=18'} + cpu: [x64] + os: [win32] + '@eslint-community/eslint-utils@4.9.1': resolution: {integrity: sha512-phrYmNiYppR7znFEdqgfWHXR6NCkZEK7hwWDHZUjit/2/U0r6XvkDl0SYnoM51Hq7FhCGdLDT6zxCCOY1hexsQ==} engines: {node: ^12.22.0 || ^14.17.0 || >=16.0.0} @@ -212,6 +371,11 @@ packages: deep-is@0.1.4: resolution: {integrity: sha512-oIPzksmTg4/MriiaYGO+okXDT7ztn/w3Eptv/+gSIdMdKsJo0u4CfYNFJPy+4SKMuCqGw2wxnA+URMg3t8a/bQ==} + esbuild@0.28.2: + resolution: {integrity: sha512-HKVLS8dvII+xoKW9kmqxbRKrnWEXfJJr/FZhhJmiqIB0e053QNYFqOBouTMO/k5sID4MvCiUCvv8b9M4h32wIA==} + engines: {node: '>=18'} + hasBin: true + escape-string-regexp@4.0.0: resolution: {integrity: sha512-TtpcNJ3XAzx3Gq8sWRzJaVajRs0uVxA2YAkdb1jm2YkPz4G6egUFAyA3n5vtEIZefPk5Wa4UXbKuS5fKkJWdgA==} engines: {node: '>=10'} @@ -456,6 +620,84 @@ packages: snapshots: + '@esbuild/aix-ppc64@0.28.2': + optional: true + + '@esbuild/android-arm64@0.28.2': + optional: true + + '@esbuild/android-arm@0.28.2': + optional: true + + '@esbuild/android-x64@0.28.2': + optional: true + + '@esbuild/darwin-arm64@0.28.2': + optional: true + + '@esbuild/darwin-x64@0.28.2': + optional: true + + '@esbuild/freebsd-arm64@0.28.2': + optional: true + + '@esbuild/freebsd-x64@0.28.2': + optional: true + + '@esbuild/linux-arm64@0.28.2': + optional: true + + '@esbuild/linux-arm@0.28.2': + optional: true + + '@esbuild/linux-ia32@0.28.2': + optional: true + + '@esbuild/linux-loong64@0.28.2': + optional: true + + '@esbuild/linux-mips64el@0.28.2': + optional: true + + '@esbuild/linux-ppc64@0.28.2': + optional: true + + '@esbuild/linux-riscv64@0.28.2': + optional: true + + '@esbuild/linux-s390x@0.28.2': + optional: true + + '@esbuild/linux-x64@0.28.2': + optional: true + + '@esbuild/netbsd-arm64@0.28.2': + optional: true + + '@esbuild/netbsd-x64@0.28.2': + optional: true + + '@esbuild/openbsd-arm64@0.28.2': + optional: true + + '@esbuild/openbsd-x64@0.28.2': + optional: true + + '@esbuild/openharmony-arm64@0.28.2': + optional: true + + '@esbuild/sunos-x64@0.28.2': + optional: true + + '@esbuild/win32-arm64@0.28.2': + optional: true + + '@esbuild/win32-ia32@0.28.2': + optional: true + + '@esbuild/win32-x64@0.28.2': + optional: true + '@eslint-community/eslint-utils@4.9.1(eslint@10.1.0)': dependencies: eslint: 10.1.0 @@ -646,6 +888,35 @@ snapshots: deep-is@0.1.4: {} + esbuild@0.28.2: + optionalDependencies: + '@esbuild/aix-ppc64': 0.28.2 + '@esbuild/android-arm': 0.28.2 + '@esbuild/android-arm64': 0.28.2 + '@esbuild/android-x64': 0.28.2 + '@esbuild/darwin-arm64': 0.28.2 + '@esbuild/darwin-x64': 0.28.2 + '@esbuild/freebsd-arm64': 0.28.2 + '@esbuild/freebsd-x64': 0.28.2 + '@esbuild/linux-arm': 0.28.2 + '@esbuild/linux-arm64': 0.28.2 + '@esbuild/linux-ia32': 0.28.2 + '@esbuild/linux-loong64': 0.28.2 + '@esbuild/linux-mips64el': 0.28.2 + '@esbuild/linux-ppc64': 0.28.2 + '@esbuild/linux-riscv64': 0.28.2 + '@esbuild/linux-s390x': 0.28.2 + '@esbuild/linux-x64': 0.28.2 + '@esbuild/netbsd-arm64': 0.28.2 + '@esbuild/netbsd-x64': 0.28.2 + '@esbuild/openbsd-arm64': 0.28.2 + '@esbuild/openbsd-x64': 0.28.2 + '@esbuild/openharmony-arm64': 0.28.2 + '@esbuild/sunos-x64': 0.28.2 + '@esbuild/win32-arm64': 0.28.2 + '@esbuild/win32-ia32': 0.28.2 + '@esbuild/win32-x64': 0.28.2 + escape-string-regexp@4.0.0: {} eslint-scope@9.1.2: diff --git a/pnpm-workspace.yaml b/pnpm-workspace.yaml index dbfcdbf..f778b49 100644 --- a/pnpm-workspace.yaml +++ b/pnpm-workspace.yaml @@ -3,3 +3,6 @@ packages: overrides: brace-expansion: 5.0.9 + +onlyBuiltDependencies: + - esbuild diff --git a/scripts/bootstrap-macos-local.sh b/scripts/bootstrap-macos-local.sh index 31f17d0..5c641f5 100755 --- a/scripts/bootstrap-macos-local.sh +++ b/scripts/bootstrap-macos-local.sh @@ -101,14 +101,19 @@ fi if ! pgrep -x "ollama" >/dev/null 2>&1; then echo "Starting Ollama service..." nohup ollama serve >/tmp/selectpilot-ollama.log 2>&1 & - sleep 3 fi -if pgrep -x "ollama" >/dev/null 2>&1; then - STATUS_OLLAMA_RUNNING="ok" -else +for _attempt in $(seq 1 15); do + if curl -sSf "http://127.0.0.1:11434/api/tags" >/dev/null 2>&1; then + STATUS_OLLAMA_RUNNING="ok" + break + fi + sleep 1 +done + +if [[ "$STATUS_OLLAMA_RUNNING" != "ok" ]]; then STATUS_OLLAMA_RUNNING="failed" - echo "Ollama service did not start as expected." >&2 + echo "Ollama API did not become reachable within 15 seconds." >&2 exit 1 fi @@ -128,10 +133,31 @@ CHROMEAI_OLLAMA_NUM_CTX="$NUM_CTX" \ "$ROOT/scripts/install-macos-local.sh" STATUS_LAUNCHAGENT="ok" -if curl -sSf "$BRIDGE_HEALTH_URL" >/dev/null; then +HEALTH_JSON="$(curl -sSf "$BRIDGE_HEALTH_URL" || true)" +if python3 - "$HEALTH_JSON" <<'PY' +import json +import sys + +try: + payload = json.loads(sys.argv[1]) +except (IndexError, json.JSONDecodeError): + raise SystemExit(1) + +ollama = payload.get("ollama") or {} +raise SystemExit(0 if ( + payload.get("ok") is True + and ollama.get("reachable") is True + and ollama.get("model_available") is True + and ollama.get("embed_model_available") is True +) else 1) +PY +then STATUS_BRIDGE_HEALTH="ok" else STATUS_BRIDGE_HEALTH="failed" + echo "Local bridge is reachable but the exact configured runtime contract is not healthy." >&2 + echo "$HEALTH_JSON" >&2 + exit 1 fi cat < list[dict[str, Any]]: def get_runtime_profile(key: str | None) -> RuntimeProfile: - if key and key in RUNTIME_PROFILES: + if key in RUNTIME_PROFILES: return RUNTIME_PROFILES[key] - return RUNTIME_PROFILES["fast"] + if key is None: + return RUNTIME_PROFILES["fast"] + raise ValueError(f"Unknown runtime profile: {key}") def build_bootstrap_commands(profile_key: str, project_root: str | Path) -> dict[str, str]: diff --git a/tests/e2e/extension-first-run-flow.spec.mjs b/tests/e2e/extension-first-run-flow.spec.mjs index 05a2a8b..9d90073 100644 --- a/tests/e2e/extension-first-run-flow.spec.mjs +++ b/tests/e2e/extension-first-run-flow.spec.mjs @@ -1,6 +1,6 @@ import { test, expect, chromium } from '@playwright/test'; -test('real extension first-run sends only the canonical local example', async () => { +test('real extension preserves privacy from selected text through rendered output', async () => { const executablePath = process.env.SELECTPILOT_CHROME_EXECUTABLE || chromium.executablePath(); const context = await chromium.launchPersistentContext(test.info().outputPath('extension-user-data'), { @@ -10,10 +10,21 @@ test('real extension first-run sends only the canonical local example', async () `--disable-extensions-except=${process.cwd()}`, `--load-extension=${process.cwd()}`, '--disable-web-security', + '--disable-background-networking', + '--disable-component-update', + '--disable-default-apps', + '--no-first-run', ], }); - let extractRequest = null; + const extractRequests = []; + const externalRequests = []; + context.on('request', (request) => { + const url = new URL(request.url()); + if ((url.protocol === 'http:' || url.protocol === 'https:') && !['127.0.0.1', 'localhost'].includes(url.hostname)) { + externalRequests.push(request.url()); + } + }); await context.route('http://127.0.0.1:8083/**', async (route) => { const request = route.request(); const path = new URL(request.url()).pathname; @@ -23,6 +34,7 @@ test('real extension first-run sends only the canonical local example', async () body: JSON.stringify(body), }); + if (path === '/profiles') { return fulfill({ profiles: [{ @@ -61,7 +73,7 @@ test('real extension first-run sends only the canonical local example', async () return fulfill({ ok: false, stream_enabled: false, active_streams: 0, event_version: '1' }); } if (path === '/extract') { - extractRequest = request.postDataJSON(); + extractRequests.push(request.postDataJSON()); return fulfill({ preset: 'action_brief', label: 'Action Brief', @@ -75,6 +87,11 @@ test('real extension first-run sends only the canonical local example', async () } return fulfill({ ok: true }); }); + await context.route('https://selectpilot.test/e2e-selection', (route) => route.fulfill({ + status: 200, + contentType: 'text/html', + body: 'Private launch plan

Maya owns the privacy review by Thursday. Publishing remains blocked until the review passes.

', + })); try { let [serviceWorker] = context.serviceWorkers(); @@ -86,7 +103,7 @@ test('real extension first-run sends only the canonical local example', async () const lockedResponse = await panelPage.evaluate(() => globalThis.chrome.runtime.sendMessage({ type: 'panel:extract_demo' })); expect(lockedResponse.error).toBe('Paid license required for deterministic extraction'); - expect(extractRequest).toBeNull(); + expect(extractRequests).toEqual([]); await panelPage.evaluate(async () => { const storage = await import(globalThis.chrome.runtime.getURL('licensing/license-storage.js')); @@ -100,8 +117,8 @@ test('real extension first-run sends only the canonical local example', async () await expect(panelPage.locator('#btn-first-run-example')).toBeVisible(); await panelPage.locator('#btn-first-run-example').click(); - await expect.poll(() => extractRequest).not.toBeNull(); - expect(extractRequest).toEqual({ + await expect.poll(() => extractRequests.length).toBe(1); + expect(extractRequests[0]).toEqual({ text: 'The launch review is Friday. Maya owns the privacy check by Thursday. Jonas will verify payment and store assets. Publishing stays blocked until both checks pass.', preset: 'action_brief', url: 'selectpilot://first-run', @@ -117,6 +134,49 @@ test('real extension first-run sends only the canonical local example', async () await panelPage.reload(); await expect(panelPage.locator('#btn-first-run-example')).toHaveCount(0); + + const sourcePage = await context.newPage(); + await sourcePage.goto('https://selectpilot.test/e2e-selection'); + await expect(sourcePage.locator('body')).toContainText('Maya owns the privacy review by Thursday'); + await sourcePage.locator('#selection').selectText(); + await expect.poll(() => sourcePage.evaluate(() => globalThis.getSelection()?.toString())).toContain('Maya owns the privacy review'); + await sourcePage.waitForTimeout(250); + await sourcePage.bringToFront(); + + const activeTabUrl = await panelPage.evaluate(async () => { + const [tab] = await globalThis.chrome.tabs.query({ active: true, currentWindow: true }); + return tab?.url || ''; + }); + expect(activeTabUrl).toBe('https://selectpilot.test/e2e-selection'); + const directSelection = await panelPage.evaluate(async () => { + const [tab] = await globalThis.chrome.tabs.query({ active: true, currentWindow: true }); + try { + return await globalThis.chrome.tabs.sendMessage(tab.id, { type: 'content:get_selection' }); + } catch (error) { + return { error: String(error) }; + } + }); + expect(directSelection).toEqual({ + text: { + text: 'Maya owns the privacy review by Thursday. Publishing remains blocked until the review passes.', + url: 'https://selectpilot.test/e2e-selection', + title: 'Private launch plan', + }, + }); + + await panelPage.locator('#btn-refresh').evaluate((button) => button.click()); + await expect(panelPage.locator('#selection-card')).toContainText('Maya owns the privacy review by Thursday'); + await expect(panelPage.locator('#btn-extract')).toBeEnabled(); + + externalRequests.length = 0; + await panelPage.locator('#btn-extract').evaluate((button) => button.click()); + await expect.poll(() => extractRequests.length).toBe(2); + expect(extractRequests[1].text).toBe('Maya owns the privacy review by Thursday. Publishing remains blocked until the review passes.'); + expect(extractRequests[1].url).toBe('https://selectpilot.test/e2e-selection'); + expect(extractRequests[1].title).toBe('Private launch plan'); + await expect(panelPage.locator('#result-title')).toHaveText('Action Brief'); + await expect(panelPage.locator('#workflow')).toContainText('Maya: privacy check by Thursday'); + expect(externalRequests).toEqual([]); } finally { await context.close(); } diff --git a/tests/server/test_runtime_profiles.py b/tests/server/test_runtime_profiles.py index 9a7418a..4cb4cfe 100644 --- a/tests/server/test_runtime_profiles.py +++ b/tests/server/test_runtime_profiles.py @@ -15,8 +15,11 @@ class RuntimeProfileTests(unittest.TestCase): - def test_unknown_profile_falls_back_to_fast(self) -> None: - profile = get_runtime_profile("missing") + def test_unknown_profile_fails_explicitly(self) -> None: + with self.assertRaisesRegex(ValueError, "Unknown runtime profile"): + get_runtime_profile("missing") + + profile = get_runtime_profile(None) self.assertEqual(profile.key, "fast") self.assertEqual(profile.generation_model, "gemma4:e2b-it-qat") self.assertEqual(profile.num_ctx, 16384) From 471205e14a74a1eece5f8a3702aecf9a9ca26f43 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 21:12:02 +0200 Subject: [PATCH 12/66] feat(presets): add canonical editable registry --- README.md | 2 + docs/extraction-presets.md | 20 ++ package.json | 4 +- panel/extraction-presets.generated.js | 272 ++++++++++++++++++++++++ panel/extraction-presets.generated.ts | 272 ++++++++++++++++++++++++ panel/extraction-presets.js | 34 +-- panel/extraction-presets.ts | 36 ++-- presets/extraction-presets.json | 108 ++++++++++ scripts/generate-extraction-presets.mjs | 47 ++++ server/extraction_presets.py | 199 ++++++----------- server/nano_server.py | 12 +- tests/panel/extraction-presets.test.mjs | 6 +- tests/server/test_extraction_presets.py | 21 +- 13 files changed, 837 insertions(+), 196 deletions(-) create mode 100644 docs/extraction-presets.md create mode 100644 panel/extraction-presets.generated.js create mode 100644 panel/extraction-presets.generated.ts create mode 100644 presets/extraction-presets.json create mode 100644 scripts/generate-extraction-presets.mjs diff --git a/README.md b/README.md index dc5ed3f..7c4391f 100644 --- a/README.md +++ b/README.md @@ -35,6 +35,7 @@ SelectPilot is built to keep the core workflow local-first and inspectable. - Generates canonical metadata (source, intent, timestamps) - Exports to adapter targets (e.g. Obsidian/Notion package formats) - Uses profile-based local runtime selection (Fast / Balanced / Advanced) +- Uses editable, closed-schema extraction presets from `presets/extraction-presets.json` --- @@ -134,6 +135,7 @@ Your hardware and selected profile determine latency/quality. - `server/` — local Python bridge and runtime endpoints - `api/` — extension-to-local-bridge client layer - `tests/` — E2E + privacy/no-leakage tests +- `presets/` — canonical editable extraction preset registry ([format](docs/extraction-presets.md)) --- diff --git a/docs/extraction-presets.md b/docs/extraction-presets.md new file mode 100644 index 0000000..5ecdb2d --- /dev/null +++ b/docs/extraction-presets.md @@ -0,0 +1,20 @@ +# Extraction preset contract + +The canonical preset registry is [`presets/extraction-presets.json`](../presets/extraction-presets.json). Edit that file, then run: + +```bash +pnpm build +pnpm test +``` + +Build validates the registry and generates the browser module. The Python bridge reads the same JSON directly. Unknown explicit preset keys fail; only an omitted key uses `default_preset`. + +Each preset requires: + +- a unique `key`, user-facing `label`, and `description`; +- an `intro_key` present in the schema; +- deterministic instructions; +- a JSON Schema with `type: object`, `additionalProperties: false`, and every property listed in `required`; +- ordered `[field, heading]` pairs for Markdown rendering. + +No preset may add network behavior, cloud fallback, retention, or an implicit schema default. A changed schema must include server and panel regression coverage and pass the real extension E2E/privacy gate. diff --git a/package.json b/package.json index 726db59..f402066 100644 --- a/package.json +++ b/package.json @@ -24,8 +24,8 @@ "runtime:rollback": "node ./scripts/runtime-rollback.mjs", "runtime:show-policy": "node ./scripts/runtime-show-policy.mjs", "sync:manifest-version": "node ./scripts/sync-manifest-version.mjs", - "build": "pnpm sync:manifest-version && tsc -p tsconfig.json && node ./scripts/build-extension.mjs", - "lint": "eslint --max-warnings=0 .", + "build": "pnpm sync:manifest-version && node ./scripts/generate-extraction-presets.mjs && tsc -p tsconfig.json && node ./scripts/build-extension.mjs", + "lint": "node ./scripts/generate-extraction-presets.mjs --check && eslint --max-warnings=0 .", "lint:manifest": "node ./scripts/lint-manifest.mjs", "test:e2e": "playwright test --config tests/e2e/playwright.config.mjs", "test": "tsc -p tsconfig.json && node --test tests/panel/*.test.mjs && python3 -m unittest discover -s tests/server -p 'test_*.py'", diff --git a/panel/extraction-presets.generated.js b/panel/extraction-presets.generated.js new file mode 100644 index 0000000..9a6a047 --- /dev/null +++ b/panel/extraction-presets.generated.js @@ -0,0 +1,272 @@ +// Generated from presets/extraction-presets.json. Do not edit directly. +export const DEFAULT_EXTRACTION_PRESET = "action_brief"; +export const EXTRACTION_PRESET_DEFINITIONS = [ + { + "key": "action_brief", + "label": "Action Brief", + "description": "Turn selected text into summary, actions, decisions, risks, and follow-ups.", + "intro_key": "summary", + "instructions": "Extract a tight action brief from the selected text.\n- summary: 1 to 2 sentences.\n- action_items: concrete next steps.\n- decisions: decisions already made.\n- risks: blockers or uncertainties.\n- follow_ups: open loops to revisit.", + "schema": { + "type": "object", + "properties": { + "summary": { + "type": "string" + }, + "action_items": { + "type": "array", + "items": { + "type": "string" + } + }, + "decisions": { + "type": "array", + "items": { + "type": "string" + } + }, + "risks": { + "type": "array", + "items": { + "type": "string" + } + }, + "follow_ups": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "summary", + "action_items", + "decisions", + "risks", + "follow_ups" + ], + "additionalProperties": false + }, + "sections": [ + [ + "action_items", + "Action Items" + ], + [ + "decisions", + "Decisions" + ], + [ + "risks", + "Risks" + ], + [ + "follow_ups", + "Follow-ups" + ] + ] + }, + { + "key": "generic_json", + "label": "Generic JSON", + "description": "Create reusable JSON with key points, entities, actions, and open questions.", + "intro_key": "summary", + "instructions": "Convert the selected text into reusable structured data.\n- summary: 1 to 2 sentences.\n- key_points: the most important facts.\n- entities: named people, products, teams, or concepts.\n- action_items: concrete tasks implied by the text.\n- questions: unresolved questions or ambiguities.", + "schema": { + "type": "object", + "properties": { + "summary": { + "type": "string" + }, + "key_points": { + "type": "array", + "items": { + "type": "string" + } + }, + "entities": { + "type": "array", + "items": { + "type": "string" + } + }, + "action_items": { + "type": "array", + "items": { + "type": "string" + } + }, + "questions": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "summary", + "key_points", + "entities", + "action_items", + "questions" + ], + "additionalProperties": false + }, + "sections": [ + [ + "key_points", + "Key Points" + ], + [ + "entities", + "Entities" + ], + [ + "action_items", + "Action Items" + ], + [ + "questions", + "Open Questions" + ] + ] + }, + { + "key": "job_brief", + "label": "Job Brief", + "description": "Extract a structured hiring brief from a job post or role description.", + "intro_key": "role", + "instructions": "Turn the selected text into a compact job brief.\n- role: the job title or role focus.\n- company_context: one sentence of team or company context.\n- requirements: hard requirements.\n- nice_to_haves: soft requirements or bonuses.\n- risks: ambiguity, role scope issues, or hiring risks.\n- keywords: search terms or core skills.", + "schema": { + "type": "object", + "properties": { + "role": { + "type": "string" + }, + "company_context": { + "type": "string" + }, + "requirements": { + "type": "array", + "items": { + "type": "string" + } + }, + "nice_to_haves": { + "type": "array", + "items": { + "type": "string" + } + }, + "risks": { + "type": "array", + "items": { + "type": "string" + } + }, + "keywords": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "role", + "company_context", + "requirements", + "nice_to_haves", + "risks", + "keywords" + ], + "additionalProperties": false + }, + "sections": [ + [ + "company_context", + "Company Context" + ], + [ + "requirements", + "Requirements" + ], + [ + "nice_to_haves", + "Nice to Haves" + ], + [ + "risks", + "Risks" + ], + [ + "keywords", + "Keywords" + ] + ] + }, + { + "key": "decision_log", + "label": "Decision Log", + "description": "Capture the decision, rationale, risks, open questions, and next steps.", + "intro_key": "decision", + "instructions": "Turn the selected text into a decision log entry.\n- decision: the choice or recommendation.\n- why: short reasons or tradeoffs.\n- risks: downsides and uncertainty.\n- open_questions: unresolved decisions.\n- next_steps: concrete follow-up steps.", + "schema": { + "type": "object", + "properties": { + "decision": { + "type": "string" + }, + "why": { + "type": "array", + "items": { + "type": "string" + } + }, + "risks": { + "type": "array", + "items": { + "type": "string" + } + }, + "open_questions": { + "type": "array", + "items": { + "type": "string" + } + }, + "next_steps": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "decision", + "why", + "risks", + "open_questions", + "next_steps" + ], + "additionalProperties": false + }, + "sections": [ + [ + "why", + "Why" + ], + [ + "risks", + "Risks" + ], + [ + "open_questions", + "Open Questions" + ], + [ + "next_steps", + "Next Steps" + ] + ] + } +]; diff --git a/panel/extraction-presets.generated.ts b/panel/extraction-presets.generated.ts new file mode 100644 index 0000000..80c60a3 --- /dev/null +++ b/panel/extraction-presets.generated.ts @@ -0,0 +1,272 @@ +// Generated from presets/extraction-presets.json. Do not edit directly. +export const DEFAULT_EXTRACTION_PRESET = "action_brief" as const; +export const EXTRACTION_PRESET_DEFINITIONS = [ + { + "key": "action_brief", + "label": "Action Brief", + "description": "Turn selected text into summary, actions, decisions, risks, and follow-ups.", + "intro_key": "summary", + "instructions": "Extract a tight action brief from the selected text.\n- summary: 1 to 2 sentences.\n- action_items: concrete next steps.\n- decisions: decisions already made.\n- risks: blockers or uncertainties.\n- follow_ups: open loops to revisit.", + "schema": { + "type": "object", + "properties": { + "summary": { + "type": "string" + }, + "action_items": { + "type": "array", + "items": { + "type": "string" + } + }, + "decisions": { + "type": "array", + "items": { + "type": "string" + } + }, + "risks": { + "type": "array", + "items": { + "type": "string" + } + }, + "follow_ups": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "summary", + "action_items", + "decisions", + "risks", + "follow_ups" + ], + "additionalProperties": false + }, + "sections": [ + [ + "action_items", + "Action Items" + ], + [ + "decisions", + "Decisions" + ], + [ + "risks", + "Risks" + ], + [ + "follow_ups", + "Follow-ups" + ] + ] + }, + { + "key": "generic_json", + "label": "Generic JSON", + "description": "Create reusable JSON with key points, entities, actions, and open questions.", + "intro_key": "summary", + "instructions": "Convert the selected text into reusable structured data.\n- summary: 1 to 2 sentences.\n- key_points: the most important facts.\n- entities: named people, products, teams, or concepts.\n- action_items: concrete tasks implied by the text.\n- questions: unresolved questions or ambiguities.", + "schema": { + "type": "object", + "properties": { + "summary": { + "type": "string" + }, + "key_points": { + "type": "array", + "items": { + "type": "string" + } + }, + "entities": { + "type": "array", + "items": { + "type": "string" + } + }, + "action_items": { + "type": "array", + "items": { + "type": "string" + } + }, + "questions": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "summary", + "key_points", + "entities", + "action_items", + "questions" + ], + "additionalProperties": false + }, + "sections": [ + [ + "key_points", + "Key Points" + ], + [ + "entities", + "Entities" + ], + [ + "action_items", + "Action Items" + ], + [ + "questions", + "Open Questions" + ] + ] + }, + { + "key": "job_brief", + "label": "Job Brief", + "description": "Extract a structured hiring brief from a job post or role description.", + "intro_key": "role", + "instructions": "Turn the selected text into a compact job brief.\n- role: the job title or role focus.\n- company_context: one sentence of team or company context.\n- requirements: hard requirements.\n- nice_to_haves: soft requirements or bonuses.\n- risks: ambiguity, role scope issues, or hiring risks.\n- keywords: search terms or core skills.", + "schema": { + "type": "object", + "properties": { + "role": { + "type": "string" + }, + "company_context": { + "type": "string" + }, + "requirements": { + "type": "array", + "items": { + "type": "string" + } + }, + "nice_to_haves": { + "type": "array", + "items": { + "type": "string" + } + }, + "risks": { + "type": "array", + "items": { + "type": "string" + } + }, + "keywords": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "role", + "company_context", + "requirements", + "nice_to_haves", + "risks", + "keywords" + ], + "additionalProperties": false + }, + "sections": [ + [ + "company_context", + "Company Context" + ], + [ + "requirements", + "Requirements" + ], + [ + "nice_to_haves", + "Nice to Haves" + ], + [ + "risks", + "Risks" + ], + [ + "keywords", + "Keywords" + ] + ] + }, + { + "key": "decision_log", + "label": "Decision Log", + "description": "Capture the decision, rationale, risks, open questions, and next steps.", + "intro_key": "decision", + "instructions": "Turn the selected text into a decision log entry.\n- decision: the choice or recommendation.\n- why: short reasons or tradeoffs.\n- risks: downsides and uncertainty.\n- open_questions: unresolved decisions.\n- next_steps: concrete follow-up steps.", + "schema": { + "type": "object", + "properties": { + "decision": { + "type": "string" + }, + "why": { + "type": "array", + "items": { + "type": "string" + } + }, + "risks": { + "type": "array", + "items": { + "type": "string" + } + }, + "open_questions": { + "type": "array", + "items": { + "type": "string" + } + }, + "next_steps": { + "type": "array", + "items": { + "type": "string" + } + } + }, + "required": [ + "decision", + "why", + "risks", + "open_questions", + "next_steps" + ], + "additionalProperties": false + }, + "sections": [ + [ + "why", + "Why" + ], + [ + "risks", + "Risks" + ], + [ + "open_questions", + "Open Questions" + ], + [ + "next_steps", + "Next Steps" + ] + ] + } +] as const; diff --git a/panel/extraction-presets.js b/panel/extraction-presets.js index e254cb5..e4a5091 100644 --- a/panel/extraction-presets.js +++ b/panel/extraction-presets.js @@ -1,25 +1,13 @@ -export const EXTRACTION_PRESETS = [ - { - key: 'action_brief', - label: 'Action Brief', - description: 'Turn selected text into summary, actions, decisions, risks, and follow-ups.' - }, - { - key: 'generic_json', - label: 'Generic JSON', - description: 'Create reusable JSON with key points, entities, actions, and open questions.' - }, - { - key: 'job_brief', - label: 'Job Brief', - description: 'Extract a structured hiring brief from a job post or role description.' - }, - { - key: 'decision_log', - label: 'Decision Log', - description: 'Capture the decision, rationale, risks, open questions, and next steps.' - } -]; +import { DEFAULT_EXTRACTION_PRESET, EXTRACTION_PRESET_DEFINITIONS } from './extraction-presets.generated.js'; +export const EXTRACTION_PRESETS = EXTRACTION_PRESET_DEFINITIONS.map(({ key, label, description }) => ({ + key, + label, + description, +})); export function getExtractionPreset(key) { - return EXTRACTION_PRESETS.find((preset) => preset.key === key) || EXTRACTION_PRESETS[0]; + const resolvedKey = key ?? DEFAULT_EXTRACTION_PRESET; + const preset = EXTRACTION_PRESETS.find((candidate) => candidate.key === resolvedKey); + if (!preset) + throw new Error(`Unknown extraction preset: ${resolvedKey}`); + return preset; } diff --git a/panel/extraction-presets.ts b/panel/extraction-presets.ts index e1a79fd..cd48aa2 100644 --- a/panel/extraction-presets.ts +++ b/panel/extraction-presets.ts @@ -1,4 +1,6 @@ -export type ExtractionPresetKey = 'action_brief' | 'generic_json' | 'job_brief' | 'decision_log'; +import { DEFAULT_EXTRACTION_PRESET, EXTRACTION_PRESET_DEFINITIONS } from './extraction-presets.generated.js'; + +export type ExtractionPresetKey = string; export type ExtractionPreset = { key: ExtractionPresetKey; @@ -6,29 +8,15 @@ export type ExtractionPreset = { description: string; }; -export const EXTRACTION_PRESETS: ExtractionPreset[] = [ - { - key: 'action_brief', - label: 'Action Brief', - description: 'Turn selected text into summary, actions, decisions, risks, and follow-ups.' - }, - { - key: 'generic_json', - label: 'Generic JSON', - description: 'Create reusable JSON with key points, entities, actions, and open questions.' - }, - { - key: 'job_brief', - label: 'Job Brief', - description: 'Extract a structured hiring brief from a job post or role description.' - }, - { - key: 'decision_log', - label: 'Decision Log', - description: 'Capture the decision, rationale, risks, open questions, and next steps.' - } -]; +export const EXTRACTION_PRESETS: ExtractionPreset[] = EXTRACTION_PRESET_DEFINITIONS.map(({ key, label, description }) => ({ + key, + label, + description, +})); export function getExtractionPreset(key: string | null | undefined): ExtractionPreset { - return EXTRACTION_PRESETS.find((preset) => preset.key === key) || EXTRACTION_PRESETS[0]; + const resolvedKey = key ?? DEFAULT_EXTRACTION_PRESET; + const preset = EXTRACTION_PRESETS.find((candidate) => candidate.key === resolvedKey); + if (!preset) throw new Error(`Unknown extraction preset: ${resolvedKey}`); + return preset; } diff --git a/presets/extraction-presets.json b/presets/extraction-presets.json new file mode 100644 index 0000000..ddea315 --- /dev/null +++ b/presets/extraction-presets.json @@ -0,0 +1,108 @@ +{ + "version": "1.0.0", + "default_preset": "action_brief", + "presets": [ + { + "key": "action_brief", + "label": "Action Brief", + "description": "Turn selected text into summary, actions, decisions, risks, and follow-ups.", + "intro_key": "summary", + "instructions": "Extract a tight action brief from the selected text.\n- summary: 1 to 2 sentences.\n- action_items: concrete next steps.\n- decisions: decisions already made.\n- risks: blockers or uncertainties.\n- follow_ups: open loops to revisit.", + "schema": { + "type": "object", + "properties": { + "summary": { "type": "string" }, + "action_items": { "type": "array", "items": { "type": "string" } }, + "decisions": { "type": "array", "items": { "type": "string" } }, + "risks": { "type": "array", "items": { "type": "string" } }, + "follow_ups": { "type": "array", "items": { "type": "string" } } + }, + "required": ["summary", "action_items", "decisions", "risks", "follow_ups"], + "additionalProperties": false + }, + "sections": [ + ["action_items", "Action Items"], + ["decisions", "Decisions"], + ["risks", "Risks"], + ["follow_ups", "Follow-ups"] + ] + }, + { + "key": "generic_json", + "label": "Generic JSON", + "description": "Create reusable JSON with key points, entities, actions, and open questions.", + "intro_key": "summary", + "instructions": "Convert the selected text into reusable structured data.\n- summary: 1 to 2 sentences.\n- key_points: the most important facts.\n- entities: named people, products, teams, or concepts.\n- action_items: concrete tasks implied by the text.\n- questions: unresolved questions or ambiguities.", + "schema": { + "type": "object", + "properties": { + "summary": { "type": "string" }, + "key_points": { "type": "array", "items": { "type": "string" } }, + "entities": { "type": "array", "items": { "type": "string" } }, + "action_items": { "type": "array", "items": { "type": "string" } }, + "questions": { "type": "array", "items": { "type": "string" } } + }, + "required": ["summary", "key_points", "entities", "action_items", "questions"], + "additionalProperties": false + }, + "sections": [ + ["key_points", "Key Points"], + ["entities", "Entities"], + ["action_items", "Action Items"], + ["questions", "Open Questions"] + ] + }, + { + "key": "job_brief", + "label": "Job Brief", + "description": "Extract a structured hiring brief from a job post or role description.", + "intro_key": "role", + "instructions": "Turn the selected text into a compact job brief.\n- role: the job title or role focus.\n- company_context: one sentence of team or company context.\n- requirements: hard requirements.\n- nice_to_haves: soft requirements or bonuses.\n- risks: ambiguity, role scope issues, or hiring risks.\n- keywords: search terms or core skills.", + "schema": { + "type": "object", + "properties": { + "role": { "type": "string" }, + "company_context": { "type": "string" }, + "requirements": { "type": "array", "items": { "type": "string" } }, + "nice_to_haves": { "type": "array", "items": { "type": "string" } }, + "risks": { "type": "array", "items": { "type": "string" } }, + "keywords": { "type": "array", "items": { "type": "string" } } + }, + "required": ["role", "company_context", "requirements", "nice_to_haves", "risks", "keywords"], + "additionalProperties": false + }, + "sections": [ + ["company_context", "Company Context"], + ["requirements", "Requirements"], + ["nice_to_haves", "Nice to Haves"], + ["risks", "Risks"], + ["keywords", "Keywords"] + ] + }, + { + "key": "decision_log", + "label": "Decision Log", + "description": "Capture the decision, rationale, risks, open questions, and next steps.", + "intro_key": "decision", + "instructions": "Turn the selected text into a decision log entry.\n- decision: the choice or recommendation.\n- why: short reasons or tradeoffs.\n- risks: downsides and uncertainty.\n- open_questions: unresolved decisions.\n- next_steps: concrete follow-up steps.", + "schema": { + "type": "object", + "properties": { + "decision": { "type": "string" }, + "why": { "type": "array", "items": { "type": "string" } }, + "risks": { "type": "array", "items": { "type": "string" } }, + "open_questions": { "type": "array", "items": { "type": "string" } }, + "next_steps": { "type": "array", "items": { "type": "string" } } + }, + "required": ["decision", "why", "risks", "open_questions", "next_steps"], + "additionalProperties": false + }, + "sections": [ + ["why", "Why"], + ["risks", "Risks"], + ["open_questions", "Open Questions"], + ["next_steps", "Next Steps"] + ] + } + ] +} diff --git a/scripts/generate-extraction-presets.mjs b/scripts/generate-extraction-presets.mjs new file mode 100644 index 0000000..cdad988 --- /dev/null +++ b/scripts/generate-extraction-presets.mjs @@ -0,0 +1,47 @@ +#!/usr/bin/env node + +import fs from 'node:fs/promises'; +import path from 'node:path'; +import { fileURLToPath } from 'node:url'; + +const projectRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); +const sourcePath = path.join(projectRoot, 'presets/extraction-presets.json'); +const targetPath = path.join(projectRoot, 'panel/extraction-presets.generated.ts'); + +const source = JSON.parse(await fs.readFile(sourcePath, 'utf8')); +if (!Array.isArray(source.presets) || source.presets.length === 0) { + throw new Error('Preset registry must contain at least one preset'); +} + +const keys = new Set(); +for (const preset of source.presets) { + if (!preset || typeof preset !== 'object' || !preset.key || !preset.label || !preset.description) { + throw new Error('Every preset requires key, label, and description'); + } + if (keys.has(preset.key)) throw new Error(`Duplicate preset key: ${preset.key}`); + keys.add(preset.key); + if (preset.schema?.type !== 'object' || preset.schema?.additionalProperties !== false) { + throw new Error(`Preset ${preset.key} must use a closed object schema`); + } + const propertyKeys = Object.keys(preset.schema.properties || {}); + const required = preset.schema.required || []; + if (!propertyKeys.length || propertyKeys.some((key) => !required.includes(key)) || required.some((key) => !propertyKeys.includes(key))) { + throw new Error(`Preset ${preset.key} must require every declared property exactly`); + } + if (!propertyKeys.includes(preset.intro_key)) { + throw new Error(`Preset ${preset.key} intro_key must reference a schema property`); + } +} +if (!keys.has(source.default_preset)) { + throw new Error(`Unknown default preset: ${source.default_preset}`); +} + +const output = `// Generated from presets/extraction-presets.json. Do not edit directly.\n` + + `export const DEFAULT_EXTRACTION_PRESET = ${JSON.stringify(source.default_preset)} as const;\n` + + `export const EXTRACTION_PRESET_DEFINITIONS = ${JSON.stringify(source.presets, null, 2)} as const;\n`; +if (process.argv.includes('--check')) { + const current = await fs.readFile(targetPath, 'utf8').catch(() => ''); + if (current !== output) throw new Error('Generated extraction preset module is stale; run pnpm build'); +} else { + await fs.writeFile(targetPath, output, 'utf8'); +} diff --git a/server/extraction_presets.py b/server/extraction_presets.py index 1390502..e068735 100644 --- a/server/extraction_presets.py +++ b/server/extraction_presets.py @@ -1,9 +1,14 @@ from __future__ import annotations +import json from dataclasses import dataclass +from pathlib import Path from typing import Any +PRESET_REGISTRY_PATH = Path(__file__).resolve().parents[1] / "presets" / "extraction-presets.json" + + @dataclass(frozen=True) class ExtractionPreset: key: str @@ -15,151 +20,67 @@ class ExtractionPreset: sections: tuple[tuple[str, str], ...] -EXTRACTION_PRESETS: dict[str, ExtractionPreset] = { - "action_brief": ExtractionPreset( - key="action_brief", - label="Action Brief", - description="Turn selected text into a concise action-oriented brief.", - intro_key="summary", - schema={ - "type": "object", - "properties": { - "summary": {"type": "string"}, - "action_items": {"type": "array", "items": {"type": "string"}}, - "decisions": {"type": "array", "items": {"type": "string"}}, - "risks": {"type": "array", "items": {"type": "string"}}, - "follow_ups": {"type": "array", "items": {"type": "string"}}, - }, - "required": ["summary", "action_items", "decisions", "risks", "follow_ups"], - "additionalProperties": False, - }, - instructions=( - "Extract a tight action brief from the selected text.\n" - "- summary: 1 to 2 sentences.\n" - "- action_items: concrete next steps.\n" - "- decisions: decisions already made.\n" - "- risks: blockers or uncertainties.\n" - "- follow_ups: open loops to revisit." - ), - sections=( - ("action_items", "Action Items"), - ("decisions", "Decisions"), - ("risks", "Risks"), - ("follow_ups", "Follow-ups"), - ), - ), - "generic_json": ExtractionPreset( - key="generic_json", - label="Generic JSON", - description="Create reusable structured output from any selected text.", - intro_key="summary", - schema={ - "type": "object", - "properties": { - "summary": {"type": "string"}, - "key_points": {"type": "array", "items": {"type": "string"}}, - "entities": {"type": "array", "items": {"type": "string"}}, - "action_items": {"type": "array", "items": {"type": "string"}}, - "questions": {"type": "array", "items": {"type": "string"}}, - }, - "required": ["summary", "key_points", "entities", "action_items", "questions"], - "additionalProperties": False, - }, - instructions=( - "Convert the selected text into reusable structured data.\n" - "- summary: 1 to 2 sentences.\n" - "- key_points: the most important facts.\n" - "- entities: named people, products, teams, or concepts.\n" - "- action_items: concrete tasks implied by the text.\n" - "- questions: unresolved questions or ambiguities." - ), - sections=( - ("key_points", "Key Points"), - ("entities", "Entities"), - ("action_items", "Action Items"), - ("questions", "Open Questions"), - ), - ), - "job_brief": ExtractionPreset( - key="job_brief", - label="Job Brief", - description="Turn a role description into a structured hiring brief.", - intro_key="role", - schema={ - "type": "object", - "properties": { - "role": {"type": "string"}, - "company_context": {"type": "string"}, - "requirements": {"type": "array", "items": {"type": "string"}}, - "nice_to_haves": {"type": "array", "items": {"type": "string"}}, - "risks": {"type": "array", "items": {"type": "string"}}, - "keywords": {"type": "array", "items": {"type": "string"}}, - }, - "required": ["role", "company_context", "requirements", "nice_to_haves", "risks", "keywords"], - "additionalProperties": False, - }, - instructions=( - "Turn the selected text into a compact job brief.\n" - "- role: the job title or role focus.\n" - "- company_context: one sentence of team or company context.\n" - "- requirements: hard requirements.\n" - "- nice_to_haves: soft requirements or bonuses.\n" - "- risks: ambiguity, role scope issues, or hiring risks.\n" - "- keywords: search terms or core skills." - ), - sections=( - ("company_context", "Company Context"), - ("requirements", "Requirements"), - ("nice_to_haves", "Nice to Haves"), - ("risks", "Risks"), - ("keywords", "Keywords"), - ), - ), - "decision_log": ExtractionPreset( - key="decision_log", - label="Decision Log", - description="Capture a decision, rationale, risks, and next steps.", - intro_key="decision", - schema={ - "type": "object", - "properties": { - "decision": {"type": "string"}, - "why": {"type": "array", "items": {"type": "string"}}, - "risks": {"type": "array", "items": {"type": "string"}}, - "open_questions": {"type": "array", "items": {"type": "string"}}, - "next_steps": {"type": "array", "items": {"type": "string"}}, - }, - "required": ["decision", "why", "risks", "open_questions", "next_steps"], - "additionalProperties": False, - }, - instructions=( - "Turn the selected text into a decision log entry.\n" - "- decision: the choice or recommendation.\n" - "- why: short reasons or tradeoffs.\n" - "- risks: downsides and uncertainty.\n" - "- open_questions: unresolved decisions.\n" - "- next_steps: concrete follow-up steps." - ), - sections=( - ("why", "Why"), - ("risks", "Risks"), - ("open_questions", "Open Questions"), - ("next_steps", "Next Steps"), - ), - ), -} - -DEFAULT_EXTRACTION_PRESET = "action_brief" +def _load_registry() -> tuple[str, dict[str, ExtractionPreset]]: + try: + payload = json.loads(PRESET_REGISTRY_PATH.read_text(encoding="utf-8")) + except (OSError, json.JSONDecodeError) as exc: + raise RuntimeError(f"Cannot load extraction preset registry: {PRESET_REGISTRY_PATH}") from exc + + raw_presets = payload.get("presets") + if not isinstance(raw_presets, list) or not raw_presets: + raise RuntimeError("Extraction preset registry must contain at least one preset") + + presets: dict[str, ExtractionPreset] = {} + for raw in raw_presets: + if not isinstance(raw, dict): + raise RuntimeError("Extraction preset entries must be objects") + key = str(raw.get("key") or "").strip() + if not key or key in presets: + raise RuntimeError(f"Invalid or duplicate extraction preset key: {key}") + schema = raw.get("schema") + if not isinstance(schema, dict) or schema.get("type") != "object" or schema.get("additionalProperties") is not False: + raise RuntimeError(f"Extraction preset {key} must use a closed object schema") + properties = schema.get("properties") + required = schema.get("required") + if not isinstance(properties, dict) or not isinstance(required, list) or set(properties) != set(required): + raise RuntimeError(f"Extraction preset {key} must require every declared property exactly") + intro_key = str(raw.get("intro_key") or "") + if intro_key not in properties: + raise RuntimeError(f"Extraction preset {key} has an invalid intro_key") + raw_sections = raw.get("sections") + if not isinstance(raw_sections, list) or any(not isinstance(item, list) or len(item) != 2 for item in raw_sections): + raise RuntimeError(f"Extraction preset {key} has invalid sections") + presets[key] = ExtractionPreset( + key=key, + label=str(raw.get("label") or "").strip(), + description=str(raw.get("description") or "").strip(), + intro_key=intro_key, + schema=schema, + instructions=str(raw.get("instructions") or "").strip(), + sections=tuple((str(field), str(title)) for field, title in raw_sections), + ) + + default_key = str(payload.get("default_preset") or "") + if default_key not in presets: + raise RuntimeError(f"Unknown default extraction preset: {default_key}") + return default_key, presets + + +DEFAULT_EXTRACTION_PRESET, EXTRACTION_PRESETS = _load_registry() def get_extraction_preset(key: str | None) -> ExtractionPreset: - if key and key in EXTRACTION_PRESETS: - return EXTRACTION_PRESETS[key] - return EXTRACTION_PRESETS[DEFAULT_EXTRACTION_PRESET] + resolved_key = key if key is not None else DEFAULT_EXTRACTION_PRESET + try: + return EXTRACTION_PRESETS[resolved_key] + except KeyError as exc: + raise ValueError(f"Unknown extraction preset: {resolved_key}") from exc def render_extraction_markdown(preset: ExtractionPreset, payload: dict[str, Any]) -> str: - intro_value = str(payload.get(preset.intro_key, "")).strip() or "No summary produced." + intro_value = str(payload.get(preset.intro_key, "")).strip() + if not intro_value: + raise ValueError(f"Extraction output is missing required intro field: {preset.intro_key}") lines = [f"## {preset.label}", "", intro_value] for field, title in preset.sections: diff --git a/server/nano_server.py b/server/nano_server.py index 0def16e..93dc977 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -16,6 +16,7 @@ from urllib.parse import urlparse, parse_qs from ollama_client import OllamaClient, OllamaError +from extraction_presets import get_extraction_preset from runtime_profiles import build_bootstrap_commands, list_runtime_profiles, recommend_runtime_profile DEFAULT_PORT = 8083 @@ -175,7 +176,16 @@ def validate_summarize_payload(payload: object) -> dict: def validate_extract_payload(payload: object) -> dict: body = _expect_dict(payload) _expect_string(body, "text", required=True, allow_empty=False) - _expect_string(body, "preset", required=False, allow_empty=True) + preset = _expect_string(body, "preset", required=False, allow_empty=True) + try: + get_extraction_preset(preset or None) + except ValueError as exc: + raise ValidationError( + "unknown_extraction_preset", + str(exc), + status=422, + details={"preset": preset}, + ) from exc _expect_string(body, "title", required=False, allow_empty=True) _expect_string(body, "url", required=False, allow_empty=True) _expect_optional_dict(body, "metadata") diff --git a/tests/panel/extraction-presets.test.mjs b/tests/panel/extraction-presets.test.mjs index 6a2297b..3639103 100644 --- a/tests/panel/extraction-presets.test.mjs +++ b/tests/panel/extraction-presets.test.mjs @@ -11,8 +11,6 @@ test('exposes the execution-layer extraction presets', () => { ); }); -test('falls back to action brief for unknown presets', () => { - const preset = getExtractionPreset('unknown'); - assert.equal(preset.key, 'action_brief'); - assert.equal(preset.label, 'Action Brief'); +test('rejects unknown presets without fallback', () => { + assert.throws(() => getExtractionPreset('unknown'), /Unknown extraction preset/); }); diff --git a/tests/server/test_extraction_presets.py b/tests/server/test_extraction_presets.py index 3d306e8..cf161f5 100644 --- a/tests/server/test_extraction_presets.py +++ b/tests/server/test_extraction_presets.py @@ -10,12 +10,16 @@ sys.path.insert(0, str(SERVER_DIR)) from extraction_presets import get_extraction_preset, render_extraction_markdown # noqa: E402 +from nano_server import ValidationError, validate_extract_payload # noqa: E402 class ExtractionPresetTests(unittest.TestCase): - def test_unknown_preset_falls_back_to_action_brief(self) -> None: - preset = get_extraction_preset("does-not-exist") - self.assertEqual(preset.key, "action_brief") + def test_unknown_preset_fails_without_fallback(self) -> None: + with self.assertRaisesRegex(ValueError, "Unknown extraction preset"): + get_extraction_preset("does-not-exist") + + def test_none_uses_declared_default(self) -> None: + self.assertEqual(get_extraction_preset(None).key, "action_brief") def test_markdown_renderer_emits_sections(self) -> None: preset = get_extraction_preset("decision_log") @@ -36,6 +40,17 @@ def test_markdown_renderer_emits_sections(self) -> None: self.assertIn("### Next Steps", markdown) self.assertIn("- Verify nginx config", markdown) + def test_markdown_renderer_rejects_missing_intro(self) -> None: + preset = get_extraction_preset("decision_log") + with self.assertRaisesRegex(ValueError, "missing required intro field"): + render_extraction_markdown(preset, {"decision": "", "why": []}) + + def test_request_validation_rejects_unknown_preset(self) -> None: + with self.assertRaises(ValidationError) as caught: + validate_extract_payload({"text": "content", "preset": "missing"}) + self.assertEqual(caught.exception.code, "unknown_extraction_preset") + self.assertEqual(caught.exception.status, 422) + if __name__ == "__main__": unittest.main() From 88a6f55286c48a535e7aa58b67baec33a104dbe6 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 21:24:10 +0200 Subject: [PATCH 13/66] fix(governance): make spec evidence fail closed Verified: pnpm lint && pnpm typecheck && pnpm build && pnpm test && pnpm test:privacy && pnpm test:e2e && pnpm bench:frontier && pnpm runtime:compile-policy && pnpm runtime:validate-policy && node scripts/stress-runner.mjs --- ROADMAP.md | 57 +- agent/agent-client.js | 2 + agent/agent-client.ts | 2 + agent/agent-pipeline.js | 2 + agent/agent-pipeline.ts | 2 + agent/agent-types.ts | 2 + api/endpoints.js | 2 + api/endpoints.ts | 2 + api/nano-client.js | 3 + api/nano-client.ts | 3 + api/request.js | 2 + api/request.ts | 2 + background/background.js | 2 + background/background.ts | 2 + background/entitlement-service.js | 2 + background/entitlement-service.ts | 2 + background/tier-service.js | 2 + background/tier-service.ts | 2 + billing/billing-ui.js | 2 + billing/billing-ui.ts | 2 + billing/paddle-checkout.js | 2 + billing/paddle-checkout.ts | 2 + content/content-script.js | 2 + content/content-script.ts | 2 + content/extract-audio.js | 1 + content/extract-audio.ts | 3 + content/extract-text.js | 2 + content/extract-text.ts | 4 + content/extract-video.ts | 2 + eslint.config.js | 2 + licensing/license-storage.js | 2 + licensing/license-storage.ts | 2 + licensing/license-verifier.js | 2 + licensing/license-verifier.ts | 2 + panel/extraction-presets.js | 3 + panel/extraction-presets.ts | 3 + panel/knowledge-connectors.js | 1 + panel/knowledge-connectors.ts | 3 + panel/layout/layoutRules.js | 2 + panel/layout/layoutRules.ts | 2 + panel/layout/topologyMap.js | 2 + panel/layout/topologyMap.ts | 4 + panel/panel.js | 4 + panel/panel.ts | 4 + panel/runtime-profiles.js | 1 + panel/runtime-profiles.ts | 3 + panel/state/reportStore.js | 3 + panel/state/reportStore.ts | 5 + panel/state/runtimeStore.js | 7 + panel/state/runtimeStore.ts | 10 + popup/popup.js | 2 + popup/popup.ts | 2 + reports/aggregated_metrics.json | 3 + reports/architecture_decision_report.json | 5 +- reports/bottleneck_analysis.json | 3 + reports/determinism_report.json | 3 + reports/engine_decision_report.json | 3 + reports/frontier_decisions.json | 12 + reports/frontier_report.json | 5 +- reports/model_comparison.json | 12 + reports/module_manifest.json | 166 +- reports/phase_report.json | 20 +- reports/repo_plan.json | 2 +- reports/runtime_policy_validation.json | 4 +- reports/runtime_selection_summary.json | 4 +- reports/scenario_results.json | 1344 +++++++++++++++++ reports/simulation_report.json | 5 +- reports/spec_coverage_report.json | 163 +- .../stress/tests/T001_clean_extraction.json | 12 +- reports/stress/tests/T002_ambiguity_gate.json | 12 +- .../tests/T003_contract_enforcement.json | 12 +- .../stress/tests/T004_conflicting_input.json | 12 +- reports/stress/tests/T005_large_context.json | 12 +- .../stress/tests/T006_runtime_mismatch.json | 8 +- .../tests/T007_retrieval_pollution.json | 12 +- .../stress/tests/T008_privacy_boundary.json | 12 +- .../stress/tests/T009_parallel_execution.json | 12 +- reports/stress/tests/T010_recovery_path.json | 12 +- reports/ui_contract_validation.json | 2 +- reports/verification_report.json | 46 +- runtime/model_policy.json | 207 +-- runtime/model_registry.runtime.json | 6 +- runtime/promotion_audit.json | 65 +- scripts/benchmark-local-runtime.sh | 2 + scripts/build-extension.mjs | 2 + scripts/generate-extraction-presets.mjs | 2 + scripts/lint-manifest.mjs | 2 + scripts/runtime-policy-compiler.mjs | 2 + scripts/runtime-policy-lib.mjs | 37 +- scripts/runtime-policy-validator.mjs | 2 + scripts/runtime-rollback.mjs | 2 + scripts/runtime-show-policy.mjs | 2 + scripts/spec-compiler.mjs | 57 +- scripts/stress-runner.mjs | 7 + scripts/sync-manifest-version.mjs | 2 + scripts/ui-contract-validator.mjs | 2 + selectpilot_monolith_v3.json | 62 +- server/extraction_presets.py | 2 + server/intent/ambiguity.js | 6 + server/intent/ambiguity.ts | 8 + server/intent/intentCompiler.js | 3 + server/intent/intentCompiler.ts | 5 + server/intent/resolveIntent.js | 3 + server/intent/resolveIntent.ts | 5 + server/model/runtimeRollback.js | 3 + server/model/runtimeRollback.ts | 3 + server/model/runtimeSelectorAdapter.js | 4 + server/model/runtimeSelectorAdapter.ts | 7 + server/monero_payment_server.py | 2 + server/nano_server.py | 6 + server/ollama_client.py | 3 + server/operations/contractCompiler.js | 4 + server/operations/contractCompiler.ts | 7 + server/runSelectPilot.js | 7 + server/runSelectPilot.ts | 7 + server/runtime/liveFeedback.js | 5 + server/runtime/liveFeedback.ts | 5 + server/runtime_profiles.py | 2 + server/task/analyzer.js | 3 + server/task/analyzer.ts | 5 + shared/first-run-example.ts | 2 + shared/types/benchmarkTypes.ts | 2 + shared/types/events.ts | 2 + shared/types/intent.ts | 2 + shared/types/model.ts | 2 + shared/types/runtime.ts | 2 + shared/types/runtimePolicy.ts | 2 + shared/types/uiTopology.ts | 2 + tests/e2e/extension-first-run-flow.spec.mjs | 2 + tests/e2e/panel-no-leakage.spec.mjs | 2 + tests/e2e/panel-user-flow.spec.mjs | 2 + tests/e2e/playwright.config.mjs | 2 + tests/e2e/privacy-and-runtime.spec.mjs | 2 + tests/panel/extraction-presets.test.mjs | 2 + tests/panel/first-run-example.test.mjs | 2 + tests/panel/network-boundary.test.mjs | 2 + tests/panel/storage-encryption.test.mjs | 2 + tests/server/test_extraction_presets.py | 2 + tests/server/test_model_feedback.py | 2 + tests/server/test_ollama_client.py | 2 + tests/server/test_privacy_proof.py | 2 + tests/server/test_runtime_profiles.py | 2 + tests/server/test_validation_pipeline.py | 3 + types/node-shims.d.ts | 2 + utils/dom.js | 7 + utils/dom.ts | 7 + utils/logger.js | 5 + utils/logger.ts | 5 + utils/storage.js | 2 + utils/storage.ts | 2 + utils/time.js | 5 + utils/time.ts | 5 + 152 files changed, 2166 insertions(+), 593 deletions(-) diff --git a/ROADMAP.md b/ROADMAP.md index 5825a2d..dfb7ecc 100644 --- a/ROADMAP.md +++ b/ROADMAP.md @@ -32,8 +32,8 @@ ### 4) Onboarding-friktion är för hög -- **Status:** ✅ Löst för det lokala macOS-flödet. -- **Implementerat:** `pnpm setup:local` bygger extensionen, installerar/startar Ollama vid behov, väljer hårdvaruprofil, hämtar exakta modeller, installerar LaunchAgent och failar om exakt bridge-/modellhälsa inte kan verifieras. +- **Status:** 🟡 Delvis löst (bootstrap-script finns). +- **Åtgärd kvar:** ge ett verkligt one-command bootstrap-flöde som fungerar utan manuell felsökning. ### 5) Nginx/hosts-topologi bör vara osynlig @@ -42,15 +42,14 @@ ### 6) README saknar tydligt 5-minutersflöde högst upp -- **Status:** ✅ Löst. -- **Implementerat:** trestegs Quick Start tidigt i README, ett enda setup-kommando och copy/paste-diagnostik utan dold fallback. +- **Status:** ✅ Löst i kod och dokumentation. +- **Implementerat:** README har ett trestegs **Quick Start (5 min)** och bootstrap verifierar bridge, exakt generationmodell och embeddingmodell fail-closed. +- **Verifiering kvar:** kör flödet på avsedd slutanvändarmaskin innan full-system-DoD. ### 7) Inga end-to-end-tester för extension-flödet -- **Status:** ✅ Löst för det primära selected-text-flödet. -- **Implementerat:** Playwright-konfiguration och E2E-suite för runtime/privacy + mockad selected-text response shape. -- **Implementerat nu även:** harness-baserat sidepanel E2E-flöde som verifierar privacy-indikator + lokal fetch-trafik i panelens användarflöde. -- **Implementerat:** riktig unpacked-extension-path med injicerat content script: markera text på webbsida → background/context → sidepanel-action → renderat svar, utan harness. +- **Status:** ✅ Löst. +- **Implementerat:** Playwright laddar den riktiga unpacked extensionen och verifierar markerad text → content script → background → sidepanel action → renderat svar. --- @@ -58,12 +57,12 @@ ### 8) Inga automatiska regressionstester för privacy boundary -- **Status:** ✅ Löst för kärnflödet. +- **Status:** ✅ Löst för det automatiserade core-flödet. - **Implementerat:** - servertest för privacy proof (`tests/server/test_privacy_proof.py`) - Playwright-test som verifierar local-only privacy proof och local endpoints - Playwright-test (`tests/e2e/panel-no-leakage.spec.mjs`) som assertar lokal fetch-trafik + subtil visuell privacy proof i panel-harness -- **Implementerat:** full-extension-testet registrerar all HTTP(S)-trafik under kärnoperationen och kräver noll externa anrop. + - full-extension-test som failar vid extern HTTP(S)-trafik under core-flödet ### 9) Privacy-påstående är inte synligt verifierbart för ny användare @@ -81,8 +80,8 @@ ### 11) Presets är inte tillräckligt synliga eller utbyggbara -- **Status:** 🟡 Delvis (preset-funktionalitet finns, men tydligt användar-API saknas). -- **Åtgärd kvar:** exponera presets i redigerbar JSON/YAML + dokumentera format. +- **Status:** ✅ Löst. +- **Implementerat:** kanoniskt redigerbart JSON-register, validerad serverinläsning, genererade browserkontrakt med driftkontroll samt dokumenterat format. ### 12) Ingen tydlig upgrade-path till team/self-hosted mode @@ -107,8 +106,13 @@ ### Nästa sprint (måste först) -1. Verifiera one-command onboarding och verklig Ollama-inferens på namngiven Apple Silicon-målmaskin. -2. Samla långinput-, minnes- och latensevidens för Fast/Balanced innan slutlig modellpromotion. +1. **P1.7** Full extension E2E (markera text → sidepanel action → svar) +2. **P1.8** Strikt nätverksregressionstest för hela extension-flödet +3. **P1.6** Tydlig Quick Start (5 min) i README-toppen + +### Därefter (releasekvalitet) + +4. **P1.4** Minska onboarding-friktion ytterligare i bootstrap ### Strategisk polish @@ -126,12 +130,18 @@ ### 1) One-command installation (absolut först) - **Mål:** användaren ska klara onboarding med en enda kommando-rad. -- **Status:** Implementerat och kontraktstestat; verklig målmaskinsverifiering återstår. +- **Kvar att leverera:** + - gör `pnpm bootstrap:local` helt självbärande med robust felhantering + - automatisk kontroll av Ollama, modeller, launchd och `/health` + - tydlig slutrapport med exakt "klart/återstår" och nästa steg ### 2) "89-åring-läge" i README (ultrakort onboarding) - **Mål:** inga tekniska beslut i första flödet. -- **Status:** Trestegsflöde och copy/paste-diagnostik implementerade; användbarhetsvalidering på ren målmaskin återstår. +- **Kvar att leverera:** + - 3-stegs quick start högst upp i README (kopiera kommando → ladda extension → klart) + - separera nybörjarflöde från advanced/dev-sektioner + - lägg till "om något går fel" med copy/paste-kommandon ### 3) UI/UX-polish till premium-kvalitet @@ -144,7 +154,10 @@ ### 4) E2E-bevis för verkligt användarflöde - **Mål:** kunna verifiera funktion och privacy med reproducerbara tester. -- **Status:** ✅ Implementerat: deterministisk serverlivscykel, full extension-path och strikt extern nätverksassert. +- **Kvar att leverera:** + - deterministisk start/stop av lokal server i testflödet + - full extension-path: markera text → sidepanel action → renderat svar + - strikt nätverksregression: ingen extern trafik i kärnflödet ### 5) Strategisk finish (efter kvalitet + onboarding) @@ -162,18 +175,18 @@ | 1 | Versionsdrift | P0 | Löst | Låg | | 2 | Dubbla lock-filer | P0 | Löst | Låg | | 3 | Ogiltiga manifest-fält | P0 | Löst | Låg | -| 4 | Onboarding-friktion | P1 | Löst* | Hög | +| 4 | Onboarding-friktion | P1 | Delvis | Hög | | 5 | Nginx/hosts osynlighet | P1 | Löst* | Hög | -| 6 | README Quick Start | P1 | Löst | Låg | -| 7 | E2E-tester extension | P1 | Löst | Medium | -| 8 | Privacy regressionstester | P1 | Löst | Medium | +| 6 | README Quick Start | P1 | Delvis | Låg | +| 7 | E2E-tester extension | P1 | Delvis | Medium | +| 8 | Privacy regressionstester | P1 | Delvis | Medium | | 9 | Verifierbar privacy-yta | P1 | Löst | Medium | | 10 | Omformulera positioning | P2 | Delvis | Låg | | 11 | Utbyggbara presets | P2 | Delvis | Medium | | 12 | Team/self-hosted-path | P2 | Delvis | Låg | | 13 | Tier/pricing-integritet | P2 | Delvis | Medium | -\* Kod- och kontraktsnivå är löst; verklig ren målmaskinsverifiering återstår före fullsystems-DoD. +\* Löst i huvudflödet, men behöver fortsatt skyddas från regression i docs/scripts. --- diff --git a/agent/agent-client.js b/agent/agent-client.js index dd0911e..b3dc038 100644 --- a/agent/agent-client.js +++ b/agent/agent-client.js @@ -1,3 +1,5 @@ +// module_name: agent_agent-client_ts +// spec_ref: "execution_layer" import { agent as agentEndpoint } from '../api/nano-client.js'; export async function runAgent(prompt, context) { const payload = { prompt, context }; diff --git a/agent/agent-client.ts b/agent/agent-client.ts index 0231a89..0f6dccb 100644 --- a/agent/agent-client.ts +++ b/agent/agent-client.ts @@ -1,3 +1,5 @@ +// module_name: agent_agent-client_ts +// spec_ref: "execution_layer" import { agent as agentEndpoint } from '../api/nano-client.js'; import { AgentContext, AgentOutput } from './agent-types.js'; diff --git a/agent/agent-pipeline.js b/agent/agent-pipeline.js index d170b42..9d99b60 100644 --- a/agent/agent-pipeline.js +++ b/agent/agent-pipeline.js @@ -1,3 +1,5 @@ +// module_name: agent_agent-pipeline_ts +// spec_ref: "execution_layer" import { runAgent } from './agent-client.js'; function detectInput(context) { if (context.media?.videoFrame) diff --git a/agent/agent-pipeline.ts b/agent/agent-pipeline.ts index 23d2c13..4f87223 100644 --- a/agent/agent-pipeline.ts +++ b/agent/agent-pipeline.ts @@ -1,3 +1,5 @@ +// module_name: agent_agent-pipeline_ts +// spec_ref: "execution_layer" import { runAgent } from './agent-client.js'; import type { AgentContext, AgentOutput } from './agent-types.js'; diff --git a/agent/agent-types.ts b/agent/agent-types.ts index 8afcdcd..892cdb3 100644 --- a/agent/agent-types.ts +++ b/agent/agent-types.ts @@ -1,3 +1,5 @@ +// module_name: agent_agent-types_ts +// spec_ref: "execution_layer" export type DetectedInput = { kind: 'text' | 'audio' | 'video' | 'image'; content: string; diff --git a/api/endpoints.js b/api/endpoints.js index d950bdc..7f0539a 100644 --- a/api/endpoints.js +++ b/api/endpoints.js @@ -1,3 +1,5 @@ +// module_name: api_endpoints_ts +// spec_ref: "execution_layer" export const API_BASE = 'http://127.0.0.1:8083'; export const BILLING_BASE = 'http://127.0.0.1:8090'; export const endpoints = { diff --git a/api/endpoints.ts b/api/endpoints.ts index d3f9187..fe68cfb 100644 --- a/api/endpoints.ts +++ b/api/endpoints.ts @@ -1,3 +1,5 @@ +// module_name: api_endpoints_ts +// spec_ref: "execution_layer" export const API_BASE = 'http://127.0.0.1:8083'; export const BILLING_BASE = 'http://127.0.0.1:8090'; diff --git a/api/nano-client.js b/api/nano-client.js index e53f005..6282558 100644 --- a/api/nano-client.js +++ b/api/nano-client.js @@ -1,3 +1,5 @@ +// module_name: api_nano-client_ts +// spec_ref: "execution_layer" import { endpoints } from './endpoints.js'; import { apiRequest } from './request.js'; export async function summarize(payload) { @@ -24,6 +26,7 @@ export async function compileIntent(payload) { export async function getRuntimeMetaHealth() { return apiRequest(endpoints.runtimeMetaHealth, { method: 'GET' }); } +// @spec_ref execution_layer export function getRuntimeMetaStreamUrl(afterSeq) { if (typeof afterSeq === 'number' && Number.isFinite(afterSeq) && afterSeq > 0) { return `${endpoints.runtimeMetaStream}?after=${Math.floor(afterSeq)}`; diff --git a/api/nano-client.ts b/api/nano-client.ts index 2fe081e..9e792d2 100644 --- a/api/nano-client.ts +++ b/api/nano-client.ts @@ -1,3 +1,5 @@ +// module_name: api_nano-client_ts +// spec_ref: "execution_layer" import { endpoints } from './endpoints.js'; import { apiRequest } from './request.js'; @@ -105,6 +107,7 @@ export async function getRuntimeMetaHealth() { return apiRequest(endpoints.runtimeMetaHealth, { method: 'GET' }); } +// @spec_ref execution_layer export function getRuntimeMetaStreamUrl(afterSeq?: number) { if (typeof afterSeq === 'number' && Number.isFinite(afterSeq) && afterSeq > 0) { return `${endpoints.runtimeMetaStream}?after=${Math.floor(afterSeq)}`; diff --git a/api/request.js b/api/request.js index 84678c5..b872da0 100644 --- a/api/request.js +++ b/api/request.js @@ -1,3 +1,5 @@ +// module_name: api_request_ts +// spec_ref: "execution_layer" import { log, warn } from '../utils/logger.js'; export class ApiRequestError extends Error { status; diff --git a/api/request.ts b/api/request.ts index 1390591..4aae3f6 100644 --- a/api/request.ts +++ b/api/request.ts @@ -1,3 +1,5 @@ +// module_name: api_request_ts +// spec_ref: "execution_layer" import { log, warn } from '../utils/logger.js'; export type RequestOptions = { diff --git a/background/background.js b/background/background.js index 1f87b76..a7db2bf 100644 --- a/background/background.js +++ b/background/background.js @@ -1,3 +1,5 @@ +// module_name: background_background_ts +// spec_ref: "execution_layer" import { compileIntent, extract, summarize, transcribe, vision } from '../api/nano-client.js'; import { runPipeline } from '../agent/agent-pipeline.js'; import { log, error } from '../utils/logger.js'; diff --git a/background/background.ts b/background/background.ts index bf05d7f..068703c 100644 --- a/background/background.ts +++ b/background/background.ts @@ -1,3 +1,5 @@ +// module_name: background_background_ts +// spec_ref: "execution_layer" import { compileIntent, extract, summarize, transcribe, vision } from '../api/nano-client.js'; import { runPipeline } from '../agent/agent-pipeline.js'; import { log, error } from '../utils/logger.js'; diff --git a/background/entitlement-service.js b/background/entitlement-service.js index db42faa..da1c0c7 100644 --- a/background/entitlement-service.js +++ b/background/entitlement-service.js @@ -1,3 +1,5 @@ +// module_name: background_entitlement-service_ts +// spec_ref: "execution_layer" import { endpoints } from '../api/endpoints.js'; import { error, log, warn } from '../utils/logger.js'; import { loadLicense, loadToken, saveLicense } from '../licensing/license-storage.js'; diff --git a/background/entitlement-service.ts b/background/entitlement-service.ts index 1129a23..da2b303 100644 --- a/background/entitlement-service.ts +++ b/background/entitlement-service.ts @@ -1,3 +1,5 @@ +// module_name: background_entitlement-service_ts +// spec_ref: "execution_layer" import { endpoints } from '../api/endpoints.js'; import { error, log, warn } from '../utils/logger.js'; import { loadLicense, loadToken, saveLicense, type LicenseRecord } from '../licensing/license-storage.js'; diff --git a/background/tier-service.js b/background/tier-service.js index e6670eb..e3a776b 100644 --- a/background/tier-service.js +++ b/background/tier-service.js @@ -1,3 +1,5 @@ +// module_name: background_tier-service_ts +// spec_ref: "execution_layer" import { getEntitlementTier, hasEntitlementFeature, refreshEntitlement, setEntitlementToken, } from './entitlement-service.js'; async function loadJSON(path) { const url = chrome.runtime.getURL(path); diff --git a/background/tier-service.ts b/background/tier-service.ts index ae970d0..546bd6a 100644 --- a/background/tier-service.ts +++ b/background/tier-service.ts @@ -1,3 +1,5 @@ +// module_name: background_tier-service_ts +// spec_ref: "execution_layer" import { getEntitlementTier, hasEntitlementFeature, diff --git a/billing/billing-ui.js b/billing/billing-ui.js index 10a2328..fe2dddc 100644 --- a/billing/billing-ui.js +++ b/billing/billing-ui.js @@ -1,3 +1,5 @@ +// module_name: billing_billing-ui_ts +// spec_ref: "validation_layer" import { openCheckout } from './paddle-checkout.js'; import { getPricing } from '../background/tier-service.js'; import { button, setHTML } from '../utils/dom.js'; diff --git a/billing/billing-ui.ts b/billing/billing-ui.ts index b8b9b05..7b6861e 100644 --- a/billing/billing-ui.ts +++ b/billing/billing-ui.ts @@ -1,3 +1,5 @@ +// module_name: billing_billing-ui_ts +// spec_ref: "validation_layer" import { openCheckout } from './paddle-checkout.js'; import { getPricing } from '../background/tier-service.js'; import { button, setHTML } from '../utils/dom.js'; diff --git a/billing/paddle-checkout.js b/billing/paddle-checkout.js index c8161e2..6f26fd9 100644 --- a/billing/paddle-checkout.js +++ b/billing/paddle-checkout.js @@ -1,3 +1,5 @@ +// module_name: billing_paddle-checkout_ts +// spec_ref: "validation_layer" import { log, warn } from '../utils/logger.js'; import { getPricing } from '../background/tier-service.js'; const PADDLE_SCRIPT = 'https://cdn.paddle.com/paddle/paddle.js'; diff --git a/billing/paddle-checkout.ts b/billing/paddle-checkout.ts index 7c3d3df..d8db105 100644 --- a/billing/paddle-checkout.ts +++ b/billing/paddle-checkout.ts @@ -1,3 +1,5 @@ +// module_name: billing_paddle-checkout_ts +// spec_ref: "validation_layer" import { log, warn } from '../utils/logger.js'; import { getPricing } from '../background/tier-service.js'; diff --git a/content/content-script.js b/content/content-script.js index bd8134c..2b12dc9 100644 --- a/content/content-script.js +++ b/content/content-script.js @@ -1,3 +1,5 @@ +// module_name: content_content-script_ts +// spec_ref: "execution_layer" import { extractSelection, extractDocumentText } from './extract-text.js'; import { extractAudio } from './extract-audio.js'; import { extractVideo } from './extract-video.js'; diff --git a/content/content-script.ts b/content/content-script.ts index be69087..99428aa 100644 --- a/content/content-script.ts +++ b/content/content-script.ts @@ -1,3 +1,5 @@ +// module_name: content_content-script_ts +// spec_ref: "execution_layer" import { extractSelection, extractDocumentText } from './extract-text.js'; import { extractAudio } from './extract-audio.js'; import { extractVideo } from './extract-video.js'; diff --git a/content/extract-audio.js b/content/extract-audio.js index cd5ad1a..d2149f4 100644 --- a/content/extract-audio.js +++ b/content/extract-audio.js @@ -1,3 +1,4 @@ +// @spec_ref execution_layer export function extractAudio() { const audio = document.querySelector('audio'); if (!audio) diff --git a/content/extract-audio.ts b/content/extract-audio.ts index 4fa79ea..44d08c9 100644 --- a/content/extract-audio.ts +++ b/content/extract-audio.ts @@ -1,3 +1,5 @@ +// module_name: content_extract-audio_ts +// spec_ref: "execution_layer" export type AudioExtraction = { audioUrl?: string; duration?: number; @@ -5,6 +7,7 @@ export type AudioExtraction = { pageUrl: string; }; +// @spec_ref execution_layer export function extractAudio(): AudioExtraction | null { const audio = document.querySelector('audio'); if (!audio) return null; diff --git a/content/extract-text.js b/content/extract-text.js index 49ca8c8..4298182 100644 --- a/content/extract-text.js +++ b/content/extract-text.js @@ -1,6 +1,7 @@ function clean(text) { return text.replace(/\s+/g, ' ').trim(); } +// @spec_ref execution_layer export function extractSelection() { const selection = window.getSelection(); const text = selection ? selection.toString() : ''; @@ -10,6 +11,7 @@ export function extractSelection() { title: document.title }; } +// @spec_ref execution_layer export function extractDocumentText() { const article = document.querySelector('article'); const target = article || document.body; diff --git a/content/extract-text.ts b/content/extract-text.ts index 50d3f8e..55d3b33 100644 --- a/content/extract-text.ts +++ b/content/extract-text.ts @@ -1,3 +1,5 @@ +// module_name: content_extract-text_ts +// spec_ref: "execution_layer" export type TextExtraction = { text: string; url: string; @@ -8,6 +10,7 @@ function clean(text: string): string { return text.replace(/\s+/g, ' ').trim(); } +// @spec_ref execution_layer export function extractSelection(): TextExtraction { const selection = window.getSelection(); const text = selection ? selection.toString() : ''; @@ -18,6 +21,7 @@ export function extractSelection(): TextExtraction { }; } +// @spec_ref execution_layer export function extractDocumentText(): TextExtraction { const article = document.querySelector('article'); const target = article || document.body; diff --git a/content/extract-video.ts b/content/extract-video.ts index fd5ee59..04670fe 100644 --- a/content/extract-video.ts +++ b/content/extract-video.ts @@ -1,3 +1,5 @@ +// module_name: content_extract-video_ts +// spec_ref: "execution_layer" export type VideoExtraction = { poster?: string; currentTime?: number; diff --git a/eslint.config.js b/eslint.config.js index 7da73bc..56fed04 100644 --- a/eslint.config.js +++ b/eslint.config.js @@ -1,3 +1,5 @@ +// module_name: eslint_config_js +// spec_ref: "frontend_state_contract" import js from '@eslint/js'; import globals from 'globals'; import tseslint from 'typescript-eslint'; diff --git a/licensing/license-storage.js b/licensing/license-storage.js index ef68017..89d228f 100644 --- a/licensing/license-storage.js +++ b/licensing/license-storage.js @@ -1,3 +1,5 @@ +// module_name: licensing_license-storage_ts +// spec_ref: "validation_layer" import { getDecrypted, setEncrypted } from '../utils/storage.js'; const LICENSE_KEY = 'chromeai_license_token'; const LICENSE_META = 'chromeai_license_metadata'; diff --git a/licensing/license-storage.ts b/licensing/license-storage.ts index 1f09206..6a4cd60 100644 --- a/licensing/license-storage.ts +++ b/licensing/license-storage.ts @@ -1,3 +1,5 @@ +// module_name: licensing_license-storage_ts +// spec_ref: "validation_layer" import { getDecrypted, setEncrypted } from '../utils/storage.js'; const LICENSE_KEY = 'chromeai_license_token'; diff --git a/licensing/license-verifier.js b/licensing/license-verifier.js index a54193a..fa41da9 100644 --- a/licensing/license-verifier.js +++ b/licensing/license-verifier.js @@ -1,3 +1,5 @@ +// module_name: licensing_license-verifier_ts +// spec_ref: "validation_layer" import { endpoints } from '../api/endpoints.js'; import { apiRequest } from '../api/request.js'; import { daysFrom } from '../utils/time.js'; diff --git a/licensing/license-verifier.ts b/licensing/license-verifier.ts index d5a8460..fd30076 100644 --- a/licensing/license-verifier.ts +++ b/licensing/license-verifier.ts @@ -1,3 +1,5 @@ +// module_name: licensing_license-verifier_ts +// spec_ref: "validation_layer" import { endpoints } from '../api/endpoints.js'; import { apiRequest } from '../api/request.js'; import { daysFrom } from '../utils/time.js'; diff --git a/panel/extraction-presets.js b/panel/extraction-presets.js index e4a5091..91140c1 100644 --- a/panel/extraction-presets.js +++ b/panel/extraction-presets.js @@ -1,9 +1,12 @@ +// module_name: panel_extraction-presets_ts +// spec_ref: "frontend_state_contract" import { DEFAULT_EXTRACTION_PRESET, EXTRACTION_PRESET_DEFINITIONS } from './extraction-presets.generated.js'; export const EXTRACTION_PRESETS = EXTRACTION_PRESET_DEFINITIONS.map(({ key, label, description }) => ({ key, label, description, })); +// @spec_ref frontend_state_contract export function getExtractionPreset(key) { const resolvedKey = key ?? DEFAULT_EXTRACTION_PRESET; const preset = EXTRACTION_PRESETS.find((candidate) => candidate.key === resolvedKey); diff --git a/panel/extraction-presets.ts b/panel/extraction-presets.ts index cd48aa2..1f7fc00 100644 --- a/panel/extraction-presets.ts +++ b/panel/extraction-presets.ts @@ -1,3 +1,5 @@ +// module_name: panel_extraction-presets_ts +// spec_ref: "frontend_state_contract" import { DEFAULT_EXTRACTION_PRESET, EXTRACTION_PRESET_DEFINITIONS } from './extraction-presets.generated.js'; export type ExtractionPresetKey = string; @@ -14,6 +16,7 @@ export const EXTRACTION_PRESETS: ExtractionPreset[] = EXTRACTION_PRESET_DEFINITI description, })); +// @spec_ref frontend_state_contract export function getExtractionPreset(key: string | null | undefined): ExtractionPreset { const resolvedKey = key ?? DEFAULT_EXTRACTION_PRESET; const preset = EXTRACTION_PRESETS.find((candidate) => candidate.key === resolvedKey); diff --git a/panel/knowledge-connectors.js b/panel/knowledge-connectors.js index 39dad2c..40e2ad2 100644 --- a/panel/knowledge-connectors.js +++ b/panel/knowledge-connectors.js @@ -54,6 +54,7 @@ function toCanonicalEntries(entries) { }; }); } +// @spec_ref frontend_state_contract export function buildKnowledgePackage(target, entries) { const generatedAt = new Date().toISOString(); const canonical = toCanonicalEntries(entries); diff --git a/panel/knowledge-connectors.ts b/panel/knowledge-connectors.ts index 5080105..827cdad 100644 --- a/panel/knowledge-connectors.ts +++ b/panel/knowledge-connectors.ts @@ -1,3 +1,5 @@ +// module_name: panel_knowledge-connectors_ts +// spec_ref: "frontend_state_contract" export type KnowledgeTarget = 'generic' | 'notion' | 'craft' | 'obsidian'; export type MemoryLedgerEntry = { @@ -92,6 +94,7 @@ function toCanonicalEntries(entries: MemoryLedgerEntry[]): CanonicalEntry[] { }); } +// @spec_ref frontend_state_contract export function buildKnowledgePackage(target: KnowledgeTarget, entries: MemoryLedgerEntry[]) { const generatedAt = new Date().toISOString(); const canonical = toCanonicalEntries(entries); diff --git a/panel/layout/layoutRules.js b/panel/layout/layoutRules.js index e07d61e..2b018eb 100644 --- a/panel/layout/layoutRules.js +++ b/panel/layout/layoutRules.js @@ -1,3 +1,5 @@ +// module_name: panel_layout_layoutRules_ts +// spec_ref: "frontend_state_contract" export const BASE_GRID = 8; export const SUB_GRID = 4; export const MAX_VISIBLE_PANELS = 3; diff --git a/panel/layout/layoutRules.ts b/panel/layout/layoutRules.ts index 39c8269..4c337c2 100644 --- a/panel/layout/layoutRules.ts +++ b/panel/layout/layoutRules.ts @@ -1,3 +1,5 @@ +// module_name: panel_layout_layoutRules_ts +// spec_ref: "frontend_state_contract" export const BASE_GRID = 8; export const SUB_GRID = 4; export const MAX_VISIBLE_PANELS = 3; diff --git a/panel/layout/topologyMap.js b/panel/layout/topologyMap.js index e49d6c6..46fa83c 100644 --- a/panel/layout/topologyMap.js +++ b/panel/layout/topologyMap.js @@ -28,10 +28,12 @@ const REQUIRED_COMPONENTS = [ 'memory_shell', 'status_footer', ]; +// @spec_ref frontend_state_contract export function getTopologyForComponent(componentId) { const topology = TOPOLOGY_MAP[componentId]; return topology || null; } +// @spec_ref frontend_state_contract export function validateTopologyMap(map = TOPOLOGY_MAP) { const errors = []; for (const componentId of REQUIRED_COMPONENTS) { diff --git a/panel/layout/topologyMap.ts b/panel/layout/topologyMap.ts index b11a08e..515384e 100644 --- a/panel/layout/topologyMap.ts +++ b/panel/layout/topologyMap.ts @@ -1,3 +1,5 @@ +// module_name: panel_layout_topologyMap_ts +// spec_ref: "frontend_state_contract" import type { InformationTopology } from '../../shared/types/uiTopology.js'; export const TOPOLOGY_MAP: Record = { @@ -33,11 +35,13 @@ const REQUIRED_COMPONENTS = [ 'status_footer', ]; +// @spec_ref frontend_state_contract export function getTopologyForComponent(componentId: string): InformationTopology | null { const topology = TOPOLOGY_MAP[componentId]; return topology || null; } +// @spec_ref frontend_state_contract export function validateTopologyMap(map: Record = TOPOLOGY_MAP): { ok: boolean; errors: string[] } { const errors: string[] = []; diff --git a/panel/panel.js b/panel/panel.js index 40fbfac..288e8af 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -1,3 +1,7 @@ +// module_name: panel_surface +// spec_ref: "intent_popup" +// @spec_ref observability_layer +// @spec_ref probabilistic_suggestion_engine import { $ } from '../utils/dom.js'; import { getJSON, setJSON } from '../utils/storage.js'; import { endpoints } from '../api/endpoints.js'; diff --git a/panel/panel.ts b/panel/panel.ts index be0b875..556195e 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -1,3 +1,7 @@ +// module_name: panel_surface +// spec_ref: "intent_popup" +// @spec_ref observability_layer +// @spec_ref probabilistic_suggestion_engine import { $ } from '../utils/dom.js'; import { getJSON, setJSON } from '../utils/storage.js'; import { endpoints } from '../api/endpoints.js'; diff --git a/panel/runtime-profiles.js b/panel/runtime-profiles.js index 399b6ba..52f48ec 100644 --- a/panel/runtime-profiles.js +++ b/panel/runtime-profiles.js @@ -34,6 +34,7 @@ export const RUNTIME_PROFILES = [ command: './scripts/bootstrap-macos-local.sh --profile advanced', }, ]; +// @spec_ref frontend_state_contract export function getRuntimeProfile(key) { return RUNTIME_PROFILES.find((profile) => profile.key === key) || RUNTIME_PROFILES[0]; } diff --git a/panel/runtime-profiles.ts b/panel/runtime-profiles.ts index 65f8762..23aa38a 100644 --- a/panel/runtime-profiles.ts +++ b/panel/runtime-profiles.ts @@ -1,3 +1,5 @@ +// module_name: panel_runtime-profiles_ts +// spec_ref: "frontend_state_contract" export type RuntimeProfileKey = 'fast' | 'balanced' | 'advanced'; export type RuntimeProfile = { @@ -50,6 +52,7 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ }, ]; +// @spec_ref frontend_state_contract export function getRuntimeProfile(key: string | null | undefined): RuntimeProfile { return RUNTIME_PROFILES.find((profile) => profile.key === key) || RUNTIME_PROFILES[0]; } diff --git a/panel/state/reportStore.js b/panel/state/reportStore.js index 740df06..19d8dd8 100644 --- a/panel/state/reportStore.js +++ b/panel/state/reportStore.js @@ -3,14 +3,17 @@ const state = { determinism: null, bottleneck: null, }; +// @spec_ref frontend_state_contract export function loadFrontierReport(decisions) { state.frontier = Array.isArray(decisions) ? decisions : []; return state; } +// @spec_ref frontend_state_contract export function loadDeterminismReport(report) { state.determinism = report || null; return state; } +// @spec_ref frontend_state_contract export function loadBottleneckReport(analysis) { state.bottleneck = analysis || null; return state; diff --git a/panel/state/reportStore.ts b/panel/state/reportStore.ts index 7c16fcf..1eaf512 100644 --- a/panel/state/reportStore.ts +++ b/panel/state/reportStore.ts @@ -1,3 +1,5 @@ +// module_name: panel_state_reportStore_ts +// spec_ref: "frontend_state_contract" import type { BottleneckAnalysis, DeterminismReport, FrontierDecision } from '../../shared/types/benchmarkTypes.js'; export type ReportStoreState = { @@ -12,16 +14,19 @@ const state: ReportStoreState = { bottleneck: null, }; +// @spec_ref frontend_state_contract export function loadFrontierReport(decisions: FrontierDecision[]): ReportStoreState { state.frontier = Array.isArray(decisions) ? decisions : []; return state; } +// @spec_ref frontend_state_contract export function loadDeterminismReport(report: DeterminismReport): ReportStoreState { state.determinism = report || null; return state; } +// @spec_ref frontend_state_contract export function loadBottleneckReport(analysis: BottleneckAnalysis): ReportStoreState { state.bottleneck = analysis || null; return state; diff --git a/panel/state/runtimeStore.js b/panel/state/runtimeStore.js index c5a56a4..4a6b7e7 100644 --- a/panel/state/runtimeStore.js +++ b/panel/state/runtimeStore.js @@ -17,10 +17,12 @@ const state = { steps: [], visiblePanels: ['selection_surface', 'runtime_surface', 'report_surface'], }; +// @spec_ref frontend_state_contract export function setIntent(intent) { state.intent = String(intent || ''); return state; } +// @spec_ref frontend_state_contract export function setSelectionContext(input) { state.selectionContext = { ...state.selectionContext, @@ -28,6 +30,7 @@ export function setSelectionContext(input) { }; return state; } +// @spec_ref frontend_state_contract export function applyRuntimeEvent(event) { state.runtimeHeader = { taskFamily: event.taskFamily || state.runtimeHeader.taskFamily, @@ -45,7 +48,11 @@ export function applyRuntimeEvent(event) { } return state; } +// @spec_ref frontend_state_contract export function setVisiblePanels(panels) { state.visiblePanels = [...panels].slice(0, MAX_VISIBLE_PANELS); return state; } +// module_name: runtime_store +// spec_ref: "frontend_state_contract" +// @spec_ref runtime_event_contract diff --git a/panel/state/runtimeStore.ts b/panel/state/runtimeStore.ts index 90753e9..28c1151 100644 --- a/panel/state/runtimeStore.ts +++ b/panel/state/runtimeStore.ts @@ -1,3 +1,6 @@ +// module_name: runtime_store +// spec_ref: "frontend_state_contract" +// @spec_ref runtime_event_contract import type { VisiblePanelState } from '../../shared/types/uiTopology.js'; import type { RuntimeSelectionPath } from '../../shared/types/runtimePolicy.js'; @@ -43,11 +46,13 @@ const state: RuntimeStoreState = { visiblePanels: ['selection_surface', 'runtime_surface', 'report_surface'], }; +// @spec_ref frontend_state_contract export function setIntent(intent: string): RuntimeStoreState { state.intent = String(intent || ''); return state; } +// @spec_ref frontend_state_contract export function setSelectionContext(input: Partial): RuntimeStoreState { state.selectionContext = { ...state.selectionContext, @@ -56,6 +61,7 @@ export function setSelectionContext(input: Partial { diff --git a/popup/popup.ts b/popup/popup.ts index 60a1632..a9dcad4 100644 --- a/popup/popup.ts +++ b/popup/popup.ts @@ -1,3 +1,5 @@ +// module_name: popup_popup_ts +// spec_ref: "frontend_state_contract" import { log } from '../utils/logger.js'; const btn = document.getElementById('open-panel'); diff --git a/reports/aggregated_metrics.json b/reports/aggregated_metrics.json index 54d0050..ac1361b 100644 --- a/reports/aggregated_metrics.json +++ b/reports/aggregated_metrics.json @@ -1,4 +1,7 @@ { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "totals": { "execution_matrix_rows": 448, "runs": 4480, diff --git a/reports/architecture_decision_report.json b/reports/architecture_decision_report.json index e326597..cc8375e 100644 --- a/reports/architecture_decision_report.json +++ b/reports/architecture_decision_report.json @@ -1,5 +1,8 @@ { - "generated_at": "2026-03-26T18:08:26.167Z", + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, + "generated_at": "2026-08-18T19:23:33.758Z", "engine_recommendation": "Do not start with full C++ engine. Start with high-level orchestration and instrument everything. Add native hot paths only if measurements justify it.", "bottleneck_summary": [ "strict_json_validation", diff --git a/reports/bottleneck_analysis.json b/reports/bottleneck_analysis.json index 82e0218..c6c8ad8 100644 --- a/reports/bottleneck_analysis.json +++ b/reports/bottleneck_analysis.json @@ -1,4 +1,7 @@ { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "detected": [ "model_latency_dominance", "memory_pressure_events" diff --git a/reports/determinism_report.json b/reports/determinism_report.json index 73de3c0..0ff0150 100644 --- a/reports/determinism_report.json +++ b/reports/determinism_report.json @@ -1,4 +1,7 @@ { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "target": 0.95, "group_count": 448, "bounded_output_variance": 0.9803, diff --git a/reports/engine_decision_report.json b/reports/engine_decision_report.json index 02e4f93..46ace5c 100644 --- a/reports/engine_decision_report.json +++ b/reports/engine_decision_report.json @@ -1,4 +1,7 @@ { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "recommendation": "keep_ts_runtime", "reason": "current_costs_within_thresholds", "thresholds": { diff --git a/reports/frontier_decisions.json b/reports/frontier_decisions.json index 433b882..b75ac04 100644 --- a/reports/frontier_decisions.json +++ b/reports/frontier_decisions.json @@ -1,5 +1,8 @@ [ { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "baseline_model": "qwen2:1.5b", "candidate_model": "qwen2:0.5b", "decision": "promote", @@ -10,6 +13,9 @@ "weighted_score": 0.8854 }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "baseline_model": "qwen2:1.5b", "candidate_model": "qwen2:1.5b", "decision": "reject", @@ -20,6 +26,9 @@ "weighted_score": 0.857 }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "baseline_model": "qwen2:1.5b", "candidate_model": "qwen2:7b", "decision": "reject", @@ -30,6 +39,9 @@ "weighted_score": 0.7605 }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "baseline_model": "qwen2:1.5b", "candidate_model": "qwen2:14b", "decision": "reject", diff --git a/reports/frontier_report.json b/reports/frontier_report.json index 5157a9a..2d46a04 100644 --- a/reports/frontier_report.json +++ b/reports/frontier_report.json @@ -1,5 +1,8 @@ { - "generated_at": "2026-03-26T18:08:26.167Z", + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, + "generated_at": "2026-08-18T19:23:33.758Z", "current_frontier": [], "candidate_comparisons": [], "promotion_decisions": [], diff --git a/reports/model_comparison.json b/reports/model_comparison.json index 430e945..2de550e 100644 --- a/reports/model_comparison.json +++ b/reports/model_comparison.json @@ -1,5 +1,8 @@ [ { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "model": "qwen2:0.5b", "weighted_score": 0.8854, "schema_validity_rate": 0.9717, @@ -7,6 +10,9 @@ "p95_latency_ms": 628 }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "model": "qwen2:1.5b", "weighted_score": 0.857, "schema_validity_rate": 0.9709, @@ -14,6 +20,9 @@ "p95_latency_ms": 825 }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "model": "qwen2:7b", "weighted_score": 0.7605, "schema_validity_rate": 0.9678, @@ -21,6 +30,9 @@ "p95_latency_ms": 1582 }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "model": "qwen2:14b", "weighted_score": 0.6719, "schema_validity_rate": 0.9648, diff --git a/reports/module_manifest.json b/reports/module_manifest.json index c740b07..64cac39 100644 --- a/reports/module_manifest.json +++ b/reports/module_manifest.json @@ -1,45 +1,39 @@ [ - { - "scope": "artifacts", - "module_name": "fixtures", - "path": "fixtures", - "spec_ref": "simulation_and_benchmarking.fixtures" - }, { "scope": "frontend", "module_name": "intent_popup_component", - "path": "panel/components/IntentPopup.tsx", + "path": "panel/panel.ts", "spec_ref": "intent_popup" }, { "scope": "frontend", "module_name": "runtime_overlay_component", - "path": "panel/components/RuntimeOverlay.tsx", + "path": "panel/panel.ts", "spec_ref": "observability_layer" }, { "scope": "frontend", "module_name": "runtime_step_list_component", - "path": "panel/components/RuntimeStepList.tsx", + "path": "panel/panel.ts", "spec_ref": "observability_layer" }, { "scope": "frontend", "module_name": "suggestion_list_component", - "path": "panel/components/SuggestionList.tsx", + "path": "panel/panel.ts", "spec_ref": "probabilistic_suggestion_engine" }, { "scope": "frontend", - "module_name": "overlay_event_hook", - "path": "panel/hooks/useRuntimeEvents.ts", - "spec_ref": "runtime_event_contract" + "module_name": "runtime_store", + "path": "panel/state/runtimeStore.ts", + "spec_ref": "frontend_state_contract" }, { "scope": "frontend", - "module_name": "runtime_store", + "module_name": "overlay_event_hook", "path": "panel/state/runtimeStore.ts", - "spec_ref": "frontend_state_contract" + "spec_ref": "runtime_event_contract" }, { "scope": "artifacts", @@ -47,17 +41,29 @@ "path": "reports", "spec_ref": "reporting" }, + { + "scope": "tests", + "module_name": "benchmarks", + "path": "scripts/benchmark-local-runtime.sh", + "spec_ref": "simulation_and_benchmarking" + }, { "scope": "backend", - "module_name": "pipeline_executor", - "path": "server/execution/pipeline.ts", - "spec_ref": "execution_layer" + "module_name": "frontier_evaluator", + "path": "scripts/runtime-policy-lib.mjs", + "spec_ref": "dynamic_model_evaluation_layer" }, { "scope": "backend", - "module_name": "hardware_detector", - "path": "server/hardware/detector.ts", - "spec_ref": "hardware_detection_layer" + "module_name": "metrics_collector", + "path": "scripts/stress-runner.mjs", + "spec_ref": "simulation_and_benchmarking.metrics" + }, + { + "scope": "tests", + "module_name": "simulation", + "path": "scripts/stress-runner.mjs", + "spec_ref": "simulation_and_benchmarking" }, { "scope": "backend", @@ -68,139 +74,133 @@ { "scope": "backend", "module_name": "intent_compiler", - "path": "server/intent/compiler.ts", + "path": "server/intent/intentCompiler.ts", "spec_ref": "intent_compilation_layer" }, { "scope": "backend", "module_name": "intent_normalizer", - "path": "server/intent/normalizer.ts", + "path": "server/intent/resolveIntent.ts", "spec_ref": "intent_model" }, { "scope": "backend", - "module_name": "frontier_evaluator", - "path": "server/model/frontier.ts", - "spec_ref": "dynamic_model_evaluation_layer" + "module_name": "model_registry", + "path": "server/model/runtimeSelectorAdapter.ts", + "spec_ref": "model_registry" }, { "scope": "backend", - "module_name": "model_hysteresis", - "path": "server/model/hysteresis.ts", - "spec_ref": "model_switch_hysteresis" + "module_name": "model_selector", + "path": "server/model/runtimeSelectorAdapter.ts", + "spec_ref": "model_selection_layer" }, { "scope": "backend", - "module_name": "model_provisioner", - "path": "server/model/provisioner.ts", - "spec_ref": "provisioning_layer" + "module_name": "model_hysteresis", + "path": "server/nano_server.py", + "spec_ref": "model_switch_hysteresis" }, { "scope": "backend", "module_name": "failure_isolation", - "path": "server/model/quarantine.ts", + "path": "server/nano_server.py", "spec_ref": "failure_isolation" }, { "scope": "backend", - "module_name": "model_registry", - "path": "server/model/registry.ts", - "spec_ref": "model_registry" + "module_name": "latency_budget_engine", + "path": "server/nano_server.py", + "spec_ref": "latency_budget" }, { "scope": "backend", - "module_name": "model_selector", - "path": "server/model/selector.ts", - "spec_ref": "model_selection_layer" + "module_name": "validator_engine", + "path": "server/nano_server.py", + "spec_ref": "validation_layer" }, { "scope": "backend", - "module_name": "operation_mapper", - "path": "server/operations/mapper.ts", - "spec_ref": "operation_mapping_layer" + "module_name": "output_enforcer", + "path": "server/nano_server.py", + "spec_ref": "output_enforcement" }, { "scope": "backend", - "module_name": "template_registry", - "path": "server/operations/templates.ts", - "spec_ref": "execution_layer" + "module_name": "model_provisioner", + "path": "server/ollama_client.py", + "spec_ref": "provisioning_layer" }, { "scope": "backend", "module_name": "prompt_registry", - "path": "server/prompts/registry.ts", + "path": "server/ollama_client.py", "spec_ref": "prompt_determinism" }, { "scope": "backend", - "module_name": "event_emitter", - "path": "server/runtime/events.ts", - "spec_ref": "runtime_event_contract" + "module_name": "operation_mapper", + "path": "server/operations/contractCompiler.ts", + "spec_ref": "operation_mapping_layer" }, { "scope": "backend", - "module_name": "latency_budget_engine", - "path": "server/runtime/latencyBudget.ts", - "spec_ref": "latency_budget" + "module_name": "template_registry", + "path": "server/operations/contractCompiler.ts", + "spec_ref": "execution_layer" }, { "scope": "backend", "module_name": "memory_guard", - "path": "server/runtime/memoryGuard.ts", + "path": "server/runSelectPilot.ts", "spec_ref": "memory_guard" }, { "scope": "backend", - "module_name": "metrics_collector", - "path": "server/runtime/metrics.ts", - "spec_ref": "simulation_and_benchmarking.metrics" + "module_name": "pipeline_executor", + "path": "server/runSelectPilot.ts", + "spec_ref": "execution_layer" }, { "scope": "backend", - "module_name": "task_analyzer", - "path": "server/task/analyzer.ts", - "spec_ref": "task_analysis_layer" + "module_name": "hardware_detector", + "path": "server/runtime_profiles.py", + "spec_ref": "hardware_detection_layer" }, { "scope": "backend", - "module_name": "output_enforcer", - "path": "server/validation/enforcement.ts", - "spec_ref": "output_enforcement" + "module_name": "event_emitter", + "path": "server/runtime/liveFeedback.ts", + "spec_ref": "runtime_event_contract" }, { "scope": "backend", - "module_name": "validator_engine", - "path": "server/validation/engine.ts", - "spec_ref": "validation_layer" - }, - { - "scope": "tests", - "module_name": "benchmarks", - "path": "tests/benchmarks", - "spec_ref": "simulation_and_benchmarking" - }, - { - "scope": "tests", - "module_name": "integration", - "path": "tests/integration", - "spec_ref": "testing_strategy.integration_tests" + "module_name": "task_analyzer", + "path": "server/task/analyzer.ts", + "spec_ref": "task_analysis_layer" }, { "scope": "tests", "module_name": "privacy", - "path": "tests/privacy", + "path": "tests/e2e", "spec_ref": "privacy_and_debug_policy" }, { - "scope": "tests", - "module_name": "simulation", - "path": "tests/simulation", - "spec_ref": "simulation_and_benchmarking" + "scope": "artifacts", + "module_name": "fixtures", + "path": "tests/e2e/panel-harness.html", + "spec_ref": "simulation_and_benchmarking.fixtures" }, { "scope": "tests", "module_name": "unit", - "path": "tests/unit", + "path": "tests/panel", "spec_ref": "testing_strategy.unit_tests" + }, + { + "scope": "tests", + "module_name": "integration", + "path": "tests/server", + "spec_ref": "testing_strategy.integration_tests" } ] diff --git a/reports/phase_report.json b/reports/phase_report.json index 24cfb88..8731dc5 100644 --- a/reports/phase_report.json +++ b/reports/phase_report.json @@ -8,7 +8,7 @@ "missing_node_report" ], "completion_score": 2, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_1_repo_plan", @@ -19,7 +19,7 @@ "module_manifest.json" ], "completion_score": 34, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_2_type_contracts", @@ -32,7 +32,7 @@ "shared/types/events.ts" ], "completion_score": 4, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_3_module_scaffolds", @@ -42,7 +42,7 @@ "scaffold_report.json" ], "completion_score": 1, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_4_core_implementation", @@ -50,7 +50,7 @@ "objective": "Implement backend deterministic core before UI polish.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_5_ui_implementation", @@ -58,7 +58,7 @@ "objective": "Implement frontend surfaces after runtime contracts are stable.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_6_verification", @@ -66,7 +66,7 @@ "objective": "Run static, behavioral, privacy, and state-machine validation.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_7_simulation", @@ -74,7 +74,7 @@ "objective": "Run realistic local simulations against multiple models, hardware profiles, and task scenarios.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_8_benchmark_analysis", @@ -82,7 +82,7 @@ "objective": "Compute metrics, frontier comparisons, regressions, and promotion eligibility.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" }, { "phase_id": "phase_9_decision_reports", @@ -90,6 +90,6 @@ "objective": "Emit actionable reports on model policy, performance, and architecture choices.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-03-26T18:08:26.167Z" + "generated_at": "2026-08-18T19:23:33.758Z" } ] diff --git a/reports/repo_plan.json b/reports/repo_plan.json index 24e867c..71d07da 100644 --- a/reports/repo_plan.json +++ b/reports/repo_plan.json @@ -1,5 +1,5 @@ { - "generated_at": "2026-03-26T18:08:26.166Z", + "generated_at": "2026-08-18T19:23:33.758Z", "input_document": "selectpilot_monolith_v3.json", "phases": [ { diff --git a/reports/runtime_policy_validation.json b/reports/runtime_policy_validation.json index ca7ee84..44e54e4 100644 --- a/reports/runtime_policy_validation.json +++ b/reports/runtime_policy_validation.json @@ -15,8 +15,8 @@ "fallback_cycle_detected", "quarantined_model_promoted" ], - "policy_version": "runtime-policy-d7833f631d57", - "checked_at_unix_ms": 1700704147799, + "policy_version": "runtime-policy-c44de2e4a8a1", + "checked_at_unix_ms": 1700798162081, "policy_path": "runtime/model_policy.json", "runtime_registry_path": "runtime/model_registry.runtime.json" } diff --git a/reports/runtime_selection_summary.json b/reports/runtime_selection_summary.json index 7a6f2d6..2f0755c 100644 --- a/reports/runtime_selection_summary.json +++ b/reports/runtime_selection_summary.json @@ -1,6 +1,6 @@ { - "generated_at_unix_ms": 1700704147799, - "policy_version": "runtime-policy-d7833f631d57", + "generated_at_unix_ms": 1700798162081, + "policy_version": "runtime-policy-c44de2e4a8a1", "defaults_count": 12, "preferred_models": [ "qwen2.5:0.5b" diff --git a/reports/scenario_results.json b/reports/scenario_results.json index 3153b1b..1c4c2c3 100644 --- a/reports/scenario_results.json +++ b/reports/scenario_results.json @@ -1,5 +1,8 @@ [ { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -17,6 +20,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -34,6 +40,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -51,6 +60,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -68,6 +80,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:7b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -85,6 +100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:7b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -102,6 +120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:14b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -119,6 +140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|low|qwen2:14b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "low", @@ -136,6 +160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -153,6 +180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -170,6 +200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -187,6 +220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -204,6 +240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:7b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -221,6 +260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:7b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -238,6 +280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:14b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -255,6 +300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium|qwen2:14b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium", @@ -272,6 +320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -289,6 +340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -306,6 +360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -323,6 +380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -340,6 +400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -357,6 +420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -374,6 +440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -391,6 +460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", @@ -408,6 +480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -425,6 +500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -442,6 +520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -459,6 +540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -476,6 +560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:7b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -493,6 +580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:7b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -510,6 +600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:14b|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -527,6 +620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_minimal|high|qwen2:14b|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "high", @@ -544,6 +640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -561,6 +660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -578,6 +680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -595,6 +700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -612,6 +720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:7b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -629,6 +740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:7b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -646,6 +760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:14b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -663,6 +780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|low|qwen2:14b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "low", @@ -680,6 +800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -697,6 +820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -714,6 +840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -731,6 +860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -748,6 +880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:7b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -765,6 +900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:7b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -782,6 +920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:14b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -799,6 +940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium|qwen2:14b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium", @@ -816,6 +960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -833,6 +980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -850,6 +1000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -867,6 +1020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -884,6 +1040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -901,6 +1060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -918,6 +1080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -935,6 +1100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "medium_gpu", @@ -952,6 +1120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:0.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -969,6 +1140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:0.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -986,6 +1160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:1.5b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -1003,6 +1180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:1.5b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -1020,6 +1200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:7b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -1037,6 +1220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:7b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -1054,6 +1240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:14b|system_selected|cold", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -1071,6 +1260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "rewrite_constrained|high|qwen2:14b|system_selected|warm", "scenario": "rewrite_constrained", "hardware_profile": "high", @@ -1088,6 +1280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:0.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "low", @@ -1105,6 +1300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:0.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "low", @@ -1122,6 +1320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:1.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "low", @@ -1139,6 +1340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:1.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "low", @@ -1156,6 +1360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:7b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "low", @@ -1173,6 +1380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:7b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "low", @@ -1190,6 +1400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:14b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "low", @@ -1207,6 +1420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|low|qwen2:14b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "low", @@ -1224,6 +1440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:0.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1241,6 +1460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:0.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1258,6 +1480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:1.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1275,6 +1500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:1.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1292,6 +1520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:7b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1309,6 +1540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:7b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1326,6 +1560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:14b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1343,6 +1580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium|qwen2:14b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium", @@ -1360,6 +1600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1377,6 +1620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1394,6 +1640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1411,6 +1660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1428,6 +1680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1445,6 +1700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1462,6 +1720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1479,6 +1740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "medium_gpu", @@ -1496,6 +1760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:0.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "high", @@ -1513,6 +1780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:0.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "high", @@ -1530,6 +1800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:1.5b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "high", @@ -1547,6 +1820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:1.5b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "high", @@ -1564,6 +1840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:7b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "high", @@ -1581,6 +1860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:7b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "high", @@ -1598,6 +1880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:14b|system_selected|cold", "scenario": "classification_binary", "hardware_profile": "high", @@ -1615,6 +1900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_binary|high|qwen2:14b|system_selected|warm", "scenario": "classification_binary", "hardware_profile": "high", @@ -1632,6 +1920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:0.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1649,6 +1940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:0.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1666,6 +1960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:1.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1683,6 +1980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:1.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1700,6 +2000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:7b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1717,6 +2020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:7b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1734,6 +2040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:14b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1751,6 +2060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|low|qwen2:14b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "low", @@ -1768,6 +2080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:0.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1785,6 +2100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:0.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1802,6 +2120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:1.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1819,6 +2140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:1.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1836,6 +2160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:7b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1853,6 +2180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:7b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1870,6 +2200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:14b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1887,6 +2220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium|qwen2:14b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium", @@ -1904,6 +2240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -1921,6 +2260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -1938,6 +2280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -1955,6 +2300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -1972,6 +2320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -1989,6 +2340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -2006,6 +2360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -2023,6 +2380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "medium_gpu", @@ -2040,6 +2400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:0.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2057,6 +2420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:0.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2074,6 +2440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:1.5b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2091,6 +2460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:1.5b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2108,6 +2480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:7b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2125,6 +2500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:7b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2142,6 +2520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:14b|system_selected|cold", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2159,6 +2540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "classification_multiclass|high|qwen2:14b|system_selected|warm", "scenario": "classification_multiclass", "hardware_profile": "high", @@ -2176,6 +2560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:0.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "low", @@ -2193,6 +2580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:0.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "low", @@ -2210,6 +2600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:1.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "low", @@ -2227,6 +2620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:1.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "low", @@ -2244,6 +2640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:7b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "low", @@ -2261,6 +2660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:7b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "low", @@ -2278,6 +2680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:14b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "low", @@ -2295,6 +2700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|low|qwen2:14b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "low", @@ -2312,6 +2720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:0.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2329,6 +2740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:0.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2346,6 +2760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:1.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2363,6 +2780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:1.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2380,6 +2800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:7b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2397,6 +2820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:7b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2414,6 +2840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:14b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2431,6 +2860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium|qwen2:14b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium", @@ -2448,6 +2880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2465,6 +2900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2482,6 +2920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2499,6 +2940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2516,6 +2960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2533,6 +2980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2550,6 +3000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2567,6 +3020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "medium_gpu", @@ -2584,6 +3040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:0.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "high", @@ -2601,6 +3060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:0.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "high", @@ -2618,6 +3080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:1.5b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "high", @@ -2635,6 +3100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:1.5b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "high", @@ -2652,6 +3120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:7b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "high", @@ -2669,6 +3140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:7b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "high", @@ -2686,6 +3160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:14b|system_selected|cold", "scenario": "extract_entities", "hardware_profile": "high", @@ -2703,6 +3180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_entities|high|qwen2:14b|system_selected|warm", "scenario": "extract_entities", "hardware_profile": "high", @@ -2720,6 +3200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:0.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2737,6 +3220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:0.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2754,6 +3240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:1.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2771,6 +3260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:1.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2788,6 +3280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:7b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2805,6 +3300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:7b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2822,6 +3320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:14b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2839,6 +3340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|low|qwen2:14b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "low", @@ -2856,6 +3360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:0.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2873,6 +3380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:0.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2890,6 +3400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:1.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2907,6 +3420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:1.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2924,6 +3440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:7b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2941,6 +3460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:7b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2958,6 +3480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:14b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2975,6 +3500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium|qwen2:14b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium", @@ -2992,6 +3520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3009,6 +3540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3026,6 +3560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3043,6 +3580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3060,6 +3600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3077,6 +3620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3094,6 +3640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3111,6 +3660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", @@ -3128,6 +3680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:0.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3145,6 +3700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:0.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3162,6 +3720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:1.5b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3179,6 +3740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:1.5b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3196,6 +3760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:7b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3213,6 +3780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:7b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3230,6 +3800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:14b|system_selected|cold", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3247,6 +3820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_strict_json|high|qwen2:14b|system_selected|warm", "scenario": "extract_strict_json", "hardware_profile": "high", @@ -3264,6 +3840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:0.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3281,6 +3860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:0.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3298,6 +3880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:1.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3315,6 +3900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:1.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3332,6 +3920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:7b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3349,6 +3940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:7b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3366,6 +3960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:14b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3383,6 +3980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|low|qwen2:14b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "low", @@ -3400,6 +4000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:0.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3417,6 +4020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:0.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3434,6 +4040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:1.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3451,6 +4060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:1.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3468,6 +4080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:7b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3485,6 +4100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:7b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3502,6 +4120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:14b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3519,6 +4140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium|qwen2:14b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium", @@ -3536,6 +4160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3553,6 +4180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3570,6 +4200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3587,6 +4220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3604,6 +4240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3621,6 +4260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3638,6 +4280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3655,6 +4300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "medium_gpu", @@ -3672,6 +4320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:0.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3689,6 +4340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:0.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3706,6 +4360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:1.5b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3723,6 +4380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:1.5b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3740,6 +4400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:7b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3757,6 +4420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:7b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3774,6 +4440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:14b|system_selected|cold", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3791,6 +4460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "extract_nested_json|high|qwen2:14b|system_selected|warm", "scenario": "extract_nested_json", "hardware_profile": "high", @@ -3808,6 +4480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:0.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3825,6 +4500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:0.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3842,6 +4520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:1.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3859,6 +4540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:1.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3876,6 +4560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:7b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3893,6 +4580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:7b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3910,6 +4600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:14b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3927,6 +4620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|low|qwen2:14b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "low", @@ -3944,6 +4640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:0.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -3961,6 +4660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:0.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -3978,6 +4680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:1.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -3995,6 +4700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:1.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -4012,6 +4720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:7b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -4029,6 +4740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:7b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -4046,6 +4760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:14b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -4063,6 +4780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium|qwen2:14b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium", @@ -4080,6 +4800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4097,6 +4820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4114,6 +4840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4131,6 +4860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4148,6 +4880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4165,6 +4900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4182,6 +4920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4199,6 +4940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "medium_gpu", @@ -4216,6 +4960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:0.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4233,6 +4980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:0.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4250,6 +5000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:1.5b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4267,6 +5020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:1.5b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4284,6 +5040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:7b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4301,6 +5060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:7b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4318,6 +5080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:14b|system_selected|cold", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4335,6 +5100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "summarize_contextual|high|qwen2:14b|system_selected|warm", "scenario": "summarize_contextual", "hardware_profile": "high", @@ -4352,6 +5120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:0.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4369,6 +5140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:0.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4386,6 +5160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:1.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4403,6 +5180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:1.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4420,6 +5200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:7b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4437,6 +5220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:7b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4454,6 +5240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:14b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4471,6 +5260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|low|qwen2:14b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "low", @@ -4488,6 +5280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:0.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4505,6 +5300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:0.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4522,6 +5320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:1.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4539,6 +5340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:1.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4556,6 +5360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:7b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4573,6 +5380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:7b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4590,6 +5400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:14b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4607,6 +5420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium|qwen2:14b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium", @@ -4624,6 +5440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4641,6 +5460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4658,6 +5480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4675,6 +5500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4692,6 +5520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4709,6 +5540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4726,6 +5560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4743,6 +5580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", @@ -4760,6 +5600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:0.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4777,6 +5620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:0.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4794,6 +5640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:1.5b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4811,6 +5660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:1.5b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4828,6 +5680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:7b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4845,6 +5700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:7b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4862,6 +5720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:14b|system_selected|cold", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4879,6 +5740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "analyze_reasoning|high|qwen2:14b|system_selected|warm", "scenario": "analyze_reasoning", "hardware_profile": "high", @@ -4896,6 +5760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:0.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "low", @@ -4913,6 +5780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:0.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "low", @@ -4930,6 +5800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:1.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "low", @@ -4947,6 +5820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:1.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "low", @@ -4964,6 +5840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:7b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "low", @@ -4981,6 +5860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:7b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "low", @@ -4998,6 +5880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:14b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "low", @@ -5015,6 +5900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|low|qwen2:14b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "low", @@ -5032,6 +5920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:0.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5049,6 +5940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:0.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5066,6 +5960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:1.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5083,6 +5980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:1.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5100,6 +6000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:7b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5117,6 +6020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:7b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5134,6 +6040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:14b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5151,6 +6060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium|qwen2:14b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium", @@ -5168,6 +6080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5185,6 +6100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5202,6 +6120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5219,6 +6140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5236,6 +6160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5253,6 +6180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5270,6 +6200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5287,6 +6220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "medium_gpu", @@ -5304,6 +6240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:0.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "high", @@ -5321,6 +6260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:0.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "high", @@ -5338,6 +6280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:1.5b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "high", @@ -5355,6 +6300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:1.5b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "high", @@ -5372,6 +6320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:7b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "high", @@ -5389,6 +6340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:7b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "high", @@ -5406,6 +6360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:14b|system_selected|cold", "scenario": "compare_structured", "hardware_profile": "high", @@ -5423,6 +6380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "compare_structured|high|qwen2:14b|system_selected|warm", "scenario": "compare_structured", "hardware_profile": "high", @@ -5440,6 +6400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:0.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5457,6 +6420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:0.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5474,6 +6440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:1.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5491,6 +6460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:1.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5508,6 +6480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:7b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5525,6 +6500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:7b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5542,6 +6520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:14b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5559,6 +6540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|low|qwen2:14b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "low", @@ -5576,6 +6560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:0.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5593,6 +6580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:0.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5610,6 +6600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:1.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5627,6 +6620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:1.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5644,6 +6640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:7b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5661,6 +6660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:7b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5678,6 +6680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:14b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5695,6 +6700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium|qwen2:14b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium", @@ -5712,6 +6720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5729,6 +6740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5746,6 +6760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5763,6 +6780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5780,6 +6800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5797,6 +6820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5814,6 +6840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5831,6 +6860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "medium_gpu", @@ -5848,6 +6880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:0.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5865,6 +6900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:0.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5882,6 +6920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:1.5b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5899,6 +6940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:1.5b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5916,6 +6960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:7b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5933,6 +6980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:7b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5950,6 +7000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:14b|system_selected|cold", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5967,6 +7020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "ambiguous_intent|high|qwen2:14b|system_selected|warm", "scenario": "ambiguous_intent", "hardware_profile": "high", @@ -5984,6 +7040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:0.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6001,6 +7060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:0.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6018,6 +7080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:1.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6035,6 +7100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:1.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6052,6 +7120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:7b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6069,6 +7140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:7b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6086,6 +7160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:14b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6103,6 +7180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|low|qwen2:14b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "low", @@ -6120,6 +7200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:0.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6137,6 +7220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:0.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6154,6 +7240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:1.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6171,6 +7260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:1.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6188,6 +7280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:7b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6205,6 +7300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:7b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6222,6 +7320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:14b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6239,6 +7340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium|qwen2:14b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium", @@ -6256,6 +7360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6273,6 +7380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6290,6 +7400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6307,6 +7420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6324,6 +7440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6341,6 +7460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6358,6 +7480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6375,6 +7500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "medium_gpu", @@ -6392,6 +7520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:0.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6409,6 +7540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:0.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6426,6 +7560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:1.5b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6443,6 +7580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:1.5b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6460,6 +7600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:7b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6477,6 +7620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:7b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6494,6 +7640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:14b|system_selected|cold", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6511,6 +7660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "contradictory_intent|high|qwen2:14b|system_selected|warm", "scenario": "contradictory_intent", "hardware_profile": "high", @@ -6528,6 +7680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:0.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "low", @@ -6545,6 +7700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:0.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "low", @@ -6562,6 +7720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:1.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "low", @@ -6579,6 +7740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:1.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "low", @@ -6596,6 +7760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:7b|system_selected|cold", "scenario": "long_context", "hardware_profile": "low", @@ -6613,6 +7780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:7b|system_selected|warm", "scenario": "long_context", "hardware_profile": "low", @@ -6630,6 +7800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:14b|system_selected|cold", "scenario": "long_context", "hardware_profile": "low", @@ -6647,6 +7820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|low|qwen2:14b|system_selected|warm", "scenario": "long_context", "hardware_profile": "low", @@ -6664,6 +7840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:0.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium", @@ -6681,6 +7860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:0.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium", @@ -6698,6 +7880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:1.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium", @@ -6715,6 +7900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:1.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium", @@ -6732,6 +7920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:7b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium", @@ -6749,6 +7940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:7b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium", @@ -6766,6 +7960,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:14b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium", @@ -6783,6 +7980,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium|qwen2:14b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium", @@ -6800,6 +8000,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6817,6 +8020,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6834,6 +8040,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6851,6 +8060,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6868,6 +8080,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6885,6 +8100,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6902,6 +8120,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6919,6 +8140,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "long_context", "hardware_profile": "medium_gpu", @@ -6936,6 +8160,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:0.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "high", @@ -6953,6 +8180,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:0.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "high", @@ -6970,6 +8200,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:1.5b|system_selected|cold", "scenario": "long_context", "hardware_profile": "high", @@ -6987,6 +8220,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:1.5b|system_selected|warm", "scenario": "long_context", "hardware_profile": "high", @@ -7004,6 +8240,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:7b|system_selected|cold", "scenario": "long_context", "hardware_profile": "high", @@ -7021,6 +8260,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:7b|system_selected|warm", "scenario": "long_context", "hardware_profile": "high", @@ -7038,6 +8280,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:14b|system_selected|cold", "scenario": "long_context", "hardware_profile": "high", @@ -7055,6 +8300,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "long_context|high|qwen2:14b|system_selected|warm", "scenario": "long_context", "hardware_profile": "high", @@ -7072,6 +8320,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:0.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "low", @@ -7089,6 +8340,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:0.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "low", @@ -7106,6 +8360,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:1.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "low", @@ -7123,6 +8380,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:1.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "low", @@ -7140,6 +8400,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:7b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "low", @@ -7157,6 +8420,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:7b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "low", @@ -7174,6 +8440,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:14b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "low", @@ -7191,6 +8460,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|low|qwen2:14b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "low", @@ -7208,6 +8480,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:0.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium", @@ -7225,6 +8500,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:0.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium", @@ -7242,6 +8520,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:1.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium", @@ -7259,6 +8540,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:1.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium", @@ -7276,6 +8560,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:7b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium", @@ -7293,6 +8580,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:7b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium", @@ -7310,6 +8600,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:14b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium", @@ -7327,6 +8620,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium|qwen2:14b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium", @@ -7344,6 +8640,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:0.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7361,6 +8660,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:0.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7378,6 +8680,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:1.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7395,6 +8700,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:1.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7412,6 +8720,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:7b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7429,6 +8740,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:7b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7446,6 +8760,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:14b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7463,6 +8780,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|medium_gpu|qwen2:14b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "medium_gpu", @@ -7480,6 +8800,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:0.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "high", @@ -7497,6 +8820,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:0.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "high", @@ -7514,6 +8840,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:1.5b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "high", @@ -7531,6 +8860,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:1.5b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "high", @@ -7548,6 +8880,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:7b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "high", @@ -7565,6 +8900,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:7b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "high", @@ -7582,6 +8920,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:14b|system_selected|cold", "scenario": "noise_input", "hardware_profile": "high", @@ -7599,6 +8940,9 @@ } }, { + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, "combo_id": "noise_input|high|qwen2:14b|system_selected|warm", "scenario": "noise_input", "hardware_profile": "high", diff --git a/reports/simulation_report.json b/reports/simulation_report.json index 01025b7..4d71049 100644 --- a/reports/simulation_report.json +++ b/reports/simulation_report.json @@ -1,5 +1,8 @@ { - "generated_at": "2026-03-26T18:08:26.167Z", + "evidence_class": "deterministic_simulation", + "runtime_verified": false, + "promotion_eligible": false, + "generated_at": "2026-08-18T19:23:33.758Z", "scenario_matrix": { "dimensions": [ "task_fixture", diff --git a/reports/spec_coverage_report.json b/reports/spec_coverage_report.json index 3ce0355..fa74412 100644 --- a/reports/spec_coverage_report.json +++ b/reports/spec_coverage_report.json @@ -1,7 +1,6 @@ { - "generated_at": "2026-03-26T18:08:26.167Z", - "implemented_spec_nodes": [], - "missing_spec_nodes": [ + "generated_at": "2026-08-18T19:23:33.758Z", + "implemented_spec_nodes": [ "dynamic_model_evaluation_layer", "execution_layer", "failure_isolation", @@ -33,159 +32,7 @@ "testing_strategy.unit_tests", "validation_layer" ], - "unmapped_files": [ - "agent/agent-client.js", - "agent/agent-client.ts", - "agent/agent-pipeline.js", - "agent/agent-pipeline.ts", - "agent/agent-types.js", - "agent/agent-types.ts", - "api/endpoints.js", - "api/endpoints.ts", - "api/nano-client.js", - "api/nano-client.ts", - "api/request.js", - "api/request.ts", - "background/background.js", - "background/background.ts", - "background/entitlement-service.js", - "background/entitlement-service.ts", - "background/tier-service.js", - "background/tier-service.ts", - "billing/billing-ui.js", - "billing/billing-ui.ts", - "billing/paddle-checkout.js", - "billing/paddle-checkout.ts", - "content/content-script.js", - "content/content-script.ts", - "content/extract-audio.js", - "content/extract-audio.ts", - "content/extract-text.js", - "content/extract-text.ts", - "content/extract-video.js", - "content/extract-video.ts", - "eslint.config.js", - "licensing/license-storage.js", - "licensing/license-storage.ts", - "licensing/license-verifier.js", - "licensing/license-verifier.ts", - "panel/extraction-presets.js", - "panel/extraction-presets.ts", - "panel/knowledge-connectors.js", - "panel/knowledge-connectors.ts", - "panel/panel.js", - "panel/panel.ts", - "panel/runtime-profiles.js", - "panel/runtime-profiles.ts", - "popup/popup.js", - "popup/popup.ts", - "scripts/lint-manifest.mjs", - "scripts/spec-compiler.mjs", - "scripts/sync-manifest-version.mjs", - "server/extraction_presets.py", - "server/monero_payment_server.py", - "server/nano_server.py", - "server/ollama_client.py", - "server/runtime_profiles.py", - "shared/types/events.js", - "shared/types/events.ts", - "shared/types/intent.js", - "shared/types/intent.ts", - "shared/types/model.js", - "shared/types/model.ts", - "shared/types/runtime.js", - "shared/types/runtime.ts", - "tests/e2e/panel-no-leakage.spec.mjs", - "tests/e2e/panel-user-flow.spec.mjs", - "tests/e2e/playwright.config.mjs", - "tests/e2e/privacy-and-runtime.spec.mjs", - "tests/panel/extraction-presets.test.mjs", - "tests/panel/network-boundary.test.mjs", - "tests/server/test_extraction_presets.py", - "tests/server/test_privacy_proof.py", - "tests/server/test_runtime_profiles.py", - "tests/server/test_validation_pipeline.py", - "utils/dom.js", - "utils/dom.ts", - "utils/logger.js", - "utils/logger.ts", - "utils/storage.js", - "utils/storage.ts", - "utils/time.js", - "utils/time.ts" - ], - "unmapped_functions": [ - { - "file": "api/nano-client.js", - "missing_function_annotations": 1 - }, - { - "file": "api/nano-client.ts", - "missing_function_annotations": 1 - }, - { - "file": "content/extract-audio.js", - "missing_function_annotations": 1 - }, - { - "file": "content/extract-audio.ts", - "missing_function_annotations": 1 - }, - { - "file": "content/extract-text.js", - "missing_function_annotations": 2 - }, - { - "file": "content/extract-text.ts", - "missing_function_annotations": 2 - }, - { - "file": "panel/extraction-presets.js", - "missing_function_annotations": 1 - }, - { - "file": "panel/extraction-presets.ts", - "missing_function_annotations": 1 - }, - { - "file": "panel/knowledge-connectors.js", - "missing_function_annotations": 1 - }, - { - "file": "panel/knowledge-connectors.ts", - "missing_function_annotations": 1 - }, - { - "file": "panel/runtime-profiles.js", - "missing_function_annotations": 1 - }, - { - "file": "panel/runtime-profiles.ts", - "missing_function_annotations": 1 - }, - { - "file": "utils/dom.js", - "missing_function_annotations": 5 - }, - { - "file": "utils/dom.ts", - "missing_function_annotations": 5 - }, - { - "file": "utils/logger.js", - "missing_function_annotations": 3 - }, - { - "file": "utils/logger.ts", - "missing_function_annotations": 3 - }, - { - "file": "utils/time.js", - "missing_function_annotations": 3 - }, - { - "file": "utils/time.ts", - "missing_function_annotations": 3 - } - ] + "missing_spec_nodes": [], + "unmapped_files": [], + "unmapped_functions": [] } diff --git a/reports/stress/tests/T001_clean_extraction.json b/reports/stress/tests/T001_clean_extraction.json index ae1f526..cbf6fea 100644 --- a/reports/stress/tests/T001_clean_extraction.json +++ b/reports/stress/tests/T001_clean_extraction.json @@ -75,8 +75,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -85,12 +85,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_aaffdd1ca82d25d4", - "pipeline_trace_b": "sp_aaffdd1ca82d25d4", + "pipeline_trace_a": "sp_431776f13178edc3", + "pipeline_trace_b": "sp_431776f13178edc3", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T002_ambiguity_gate.json b/reports/stress/tests/T002_ambiguity_gate.json index 8925f03..58c9b9f 100644 --- a/reports/stress/tests/T002_ambiguity_gate.json +++ b/reports/stress/tests/T002_ambiguity_gate.json @@ -72,8 +72,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -82,12 +82,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_01ad589a15c3c028", - "pipeline_trace_b": "sp_01ad589a15c3c028", + "pipeline_trace_a": "sp_87dab4b77df5191f", + "pipeline_trace_b": "sp_87dab4b77df5191f", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T003_contract_enforcement.json b/reports/stress/tests/T003_contract_enforcement.json index a280e09..dede15a 100644 --- a/reports/stress/tests/T003_contract_enforcement.json +++ b/reports/stress/tests/T003_contract_enforcement.json @@ -69,8 +69,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -79,12 +79,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_ae1db9ac1f8ff064", - "pipeline_trace_b": "sp_ae1db9ac1f8ff064", + "pipeline_trace_a": "sp_1da3bc412de9046d", + "pipeline_trace_b": "sp_1da3bc412de9046d", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T004_conflicting_input.json b/reports/stress/tests/T004_conflicting_input.json index 983a671..404fa56 100644 --- a/reports/stress/tests/T004_conflicting_input.json +++ b/reports/stress/tests/T004_conflicting_input.json @@ -72,8 +72,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -82,12 +82,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_cfbce56462b392a6", - "pipeline_trace_b": "sp_cfbce56462b392a6", + "pipeline_trace_a": "sp_7f8aee399feeccbf", + "pipeline_trace_b": "sp_7f8aee399feeccbf", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T005_large_context.json b/reports/stress/tests/T005_large_context.json index 5094a7b..5efbc80 100644 --- a/reports/stress/tests/T005_large_context.json +++ b/reports/stress/tests/T005_large_context.json @@ -65,8 +65,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -75,12 +75,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_9a0c73a415a22dfb", - "pipeline_trace_b": "sp_9a0c73a415a22dfb", + "pipeline_trace_a": "sp_620acf79e1b90d84", + "pipeline_trace_b": "sp_620acf79e1b90d84", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T006_runtime_mismatch.json b/reports/stress/tests/T006_runtime_mismatch.json index d9a6186..0fbc506 100644 --- a/reports/stress/tests/T006_runtime_mismatch.json +++ b/reports/stress/tests/T006_runtime_mismatch.json @@ -66,7 +66,7 @@ "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_fallback", "selection_reason": "runtime_policy_fallback_no_models_available_simulated_degraded_mode", - "policy_version": "runtime-policy-d7833f631d57", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": false }, "active_model": "qwen2.5:0.5b", @@ -76,11 +76,11 @@ "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_fallback", "selection_reason": "runtime_policy_fallback_no_models_available_simulated_degraded_mode", - "policy_version": "runtime-policy-d7833f631d57", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": false }, - "pipeline_trace_a": "sp_2d9a752f83e1cc37", - "pipeline_trace_b": "sp_2d9a752f83e1cc37", + "pipeline_trace_a": "sp_eed3e21ecb6df3ca", + "pipeline_trace_b": "sp_eed3e21ecb6df3ca", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T007_retrieval_pollution.json b/reports/stress/tests/T007_retrieval_pollution.json index b050f65..271b60b 100644 --- a/reports/stress/tests/T007_retrieval_pollution.json +++ b/reports/stress/tests/T007_retrieval_pollution.json @@ -71,8 +71,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -81,12 +81,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_99f5e1f5d17a8dc2", - "pipeline_trace_b": "sp_99f5e1f5d17a8dc2", + "pipeline_trace_a": "sp_824d98e4b6fa060b", + "pipeline_trace_b": "sp_824d98e4b6fa060b", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T008_privacy_boundary.json b/reports/stress/tests/T008_privacy_boundary.json index 19a9f5e..5e7d0cd 100644 --- a/reports/stress/tests/T008_privacy_boundary.json +++ b/reports/stress/tests/T008_privacy_boundary.json @@ -71,8 +71,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -81,12 +81,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_aecadff047174d5f", - "pipeline_trace_b": "sp_aecadff047174d5f", + "pipeline_trace_a": "sp_ef61357d7ea1bae2", + "pipeline_trace_b": "sp_ef61357d7ea1bae2", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T009_parallel_execution.json b/reports/stress/tests/T009_parallel_execution.json index 207b9d4..e0c4c83 100644 --- a/reports/stress/tests/T009_parallel_execution.json +++ b/reports/stress/tests/T009_parallel_execution.json @@ -73,8 +73,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -83,12 +83,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_296a4c74be1a3da8", - "pipeline_trace_b": "sp_296a4c74be1a3da8", + "pipeline_trace_a": "sp_4b8c9d891186235f", + "pipeline_trace_b": "sp_4b8c9d891186235f", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/stress/tests/T010_recovery_path.json b/reports/stress/tests/T010_recovery_path.json index 2525e96..6515fc6 100644 --- a/reports/stress/tests/T010_recovery_path.json +++ b/reports/stress/tests/T010_recovery_path.json @@ -71,8 +71,8 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, "active_model": "qwen2.5:0.5b", @@ -81,12 +81,12 @@ "runtime_selection": { "selected_model_id": "qwen2.5:0.5b", "selection_path": "runtime_policy_preferred", - "selection_reason": "higher_correctness_same_or_lower_cost", - "policy_version": "runtime-policy-d7833f631d57", + "selection_reason": "baseline_registry_no_runtime_promotion", + "policy_version": "runtime-policy-bbc138feed76", "promotion_applied": true }, - "pipeline_trace_a": "sp_496f694144312fc1", - "pipeline_trace_b": "sp_496f694144312fc1", + "pipeline_trace_a": "sp_5c3cc2e85da13e0a", + "pipeline_trace_b": "sp_5c3cc2e85da13e0a", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, diff --git a/reports/ui_contract_validation.json b/reports/ui_contract_validation.json index d983f27..a6f7548 100644 --- a/reports/ui_contract_validation.json +++ b/reports/ui_contract_validation.json @@ -1,5 +1,5 @@ { - "generated_at": "2026-03-26T18:50:53.927Z", + "generated_at": "2026-08-18T18:41:44.263Z", "validator": "ui_contract_validator", "type": "strict_file_contract_and_mue_ui_governance", "hard_rule": "if_feature_fails_any_required_fat_dimension_it_must_not_enter_runtime_ui", diff --git a/reports/verification_report.json b/reports/verification_report.json index 471c296..06d1f5d 100644 --- a/reports/verification_report.json +++ b/reports/verification_report.json @@ -1,5 +1,8 @@ { - "generated_at": "2026-03-26T18:08:26.167Z", + "generated_at": "2026-08-18T19:23:33.758Z", + "verification_scope": "static_traceability_only", + "full_system_ok": false, + "runtime_verified": false, "static_checks": [ { "name": "typecheck", @@ -19,7 +22,7 @@ }, { "name": "spec_ref_coverage_scan", - "status": "failed" + "status": "completed" } ], "behavioral_checks": [ @@ -85,38 +88,11 @@ } ], "pass_fail_summary": { - "ok": false, - "missing_spec_refs": [ - "dynamic_model_evaluation_layer", - "execution_layer", - "failure_isolation", - "frontend_state_contract", - "hardware_detection_layer", - "intent_ambiguity_handling", - "intent_compilation_layer", - "intent_model", - "intent_popup", - "latency_budget", - "memory_guard", - "model_registry", - "model_selection_layer", - "model_switch_hysteresis", - "observability_layer", - "operation_mapping_layer", - "output_enforcement", - "privacy_and_debug_policy", - "probabilistic_suggestion_engine", - "prompt_determinism", - "provisioning_layer", - "reporting", - "runtime_event_contract", - "simulation_and_benchmarking", - "simulation_and_benchmarking.fixtures", - "simulation_and_benchmarking.metrics", - "task_analysis_layer", - "testing_strategy.integration_tests", - "testing_strategy.unit_tests", - "validation_layer" - ] + "ok": true, + "scope": "static_traceability_only", + "full_system_ok": false, + "missing_spec_refs": [], + "unmapped_files": [], + "unmapped_functions": [] } } diff --git a/runtime/model_policy.json b/runtime/model_policy.json index 7387661..2dd9426 100644 --- a/runtime/model_policy.json +++ b/runtime/model_policy.json @@ -1,6 +1,10 @@ { - "policy_version": "runtime-policy-d7833f631d57", - "generated_at_unix_ms": 1700704147799, + "policy_version": "runtime-policy-c44de2e4a8a1", + "generated_at_unix_ms": 1700798162081, + "promotion_evidence": { + "runtime_verified": false, + "status": "simulation_only_no_promotion" + }, "source_reports": [ "reports/aggregated_metrics.json", "reports/frontier_decisions.json", @@ -20,13 +24,11 @@ "output_mode": "strict_json", "preferred_model_id": "qwen2.5:0.5b", "fallback_model_ids": [], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "extract", @@ -36,13 +38,11 @@ "fallback_model_ids": [ "qwen2.5:1.5b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "extract", @@ -53,13 +53,11 @@ "qwen2.5:1.5b", "qwen2.5:7b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "extract", @@ -71,13 +69,11 @@ "qwen2.5:7b", "qwen2.5:14b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "summarize", @@ -85,13 +81,11 @@ "output_mode": "semi_structured", "preferred_model_id": "qwen2.5:0.5b", "fallback_model_ids": [], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "summarize", @@ -101,13 +95,11 @@ "fallback_model_ids": [ "qwen2.5:1.5b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "summarize", @@ -118,13 +110,11 @@ "qwen2.5:1.5b", "qwen2.5:7b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "summarize", @@ -136,13 +126,11 @@ "qwen2.5:7b", "qwen2.5:14b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "agent", @@ -150,13 +138,11 @@ "output_mode": "freeform", "preferred_model_id": "qwen2.5:0.5b", "fallback_model_ids": [], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "agent", @@ -166,13 +152,11 @@ "fallback_model_ids": [ "qwen2.5:1.5b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "agent", @@ -183,130 +167,29 @@ "qwen2.5:1.5b", "qwen2.5:7b" ], - "selection_reason": "higher_correctness_same_or_lower_cost", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700704147799 + "effective_from_unix_ms": 1700798162081 }, { "task_family": "agent", "hardware_profile": "high", "output_mode": "freeform", - "preferred_model_id": "qwen2.5:1.5b", + "preferred_model_id": "qwen2.5:0.5b", "fallback_model_ids": [ - "qwen2:0.5b", + "qwen2.5:1.5b", "qwen2.5:7b", "qwen2.5:14b" ], - "selection_reason": "rollback:manual_runtime_rollback", + "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" + "server/model/registry_source.json" ], - "effective_from_unix_ms": 1774549605424 + "effective_from_unix_ms": 1700798162081 } ], "quarantined_models": [], - "promotion_history": [ - { - "task_family": "extract", - "hardware_profile": "low", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "extract", - "hardware_profile": "medium", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "extract", - "hardware_profile": "medium_gpu", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "extract", - "hardware_profile": "high", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "summarize", - "hardware_profile": "low", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "summarize", - "hardware_profile": "medium", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "summarize", - "hardware_profile": "medium_gpu", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "summarize", - "hardware_profile": "high", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "agent", - "hardware_profile": "low", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "agent", - "hardware_profile": "medium", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "agent", - "hardware_profile": "medium_gpu", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - }, - { - "task_family": "agent", - "hardware_profile": "high", - "previous_model_id": "none", - "new_model_id": "qwen2:0.5b", - "decision_reason": "higher_correctness_same_or_lower_cost", - "effective_from_unix_ms": 1700704147799 - } - ] + "promotion_history": [] } diff --git a/runtime/model_registry.runtime.json b/runtime/model_registry.runtime.json index 6c1c3b8..15d8e40 100644 --- a/runtime/model_registry.runtime.json +++ b/runtime/model_registry.runtime.json @@ -1,6 +1,6 @@ { - "generated_at_unix_ms": 1700704147799, - "policy_version": "runtime-policy-d7833f631d57", + "generated_at_unix_ms": 1700798162081, + "policy_version": "runtime-policy-c44de2e4a8a1", "models": [ { "model_id": "qwen2.5:0.5b", @@ -28,7 +28,7 @@ ], "min_hardware_profile": "medium", "installation_state": "installed", - "runtime_status": "preferred", + "runtime_status": "fallback", "reliability_score": 0.88, "policy_refs": [ "runtime/model_policy.json#qwen2.5:1.5b" diff --git a/runtime/promotion_audit.json b/runtime/promotion_audit.json index 259a369..00ec460 100644 --- a/runtime/promotion_audit.json +++ b/runtime/promotion_audit.json @@ -1,6 +1,6 @@ { - "generated_at_unix_ms": 1700704147799, - "policy_version": "runtime-policy-d7833f631d57", + "generated_at_unix_ms": 1700798162081, + "policy_version": "runtime-policy-c44de2e4a8a1", "source_reports": [ "reports/aggregated_metrics.json", "reports/frontier_decisions.json", @@ -39,19 +39,54 @@ "reason": "manual_runtime_rollback", "task_family": "agent", "hardware_profile": "high" + }, + { + "event_type": "compile_policy", + "generated_at_unix_ms": 1700080928761, + "promoted_count": 0, + "rejected_count": 0, + "policy_version": "runtime-policy-20fa39a37df9" + }, + { + "event_type": "compile_policy", + "generated_at_unix_ms": 1700080928761, + "promoted_count": 0, + "rejected_count": 0, + "runtime_evidence_verified": false, + "policy_version": "runtime-policy-20fa39a37df9" + }, + { + "event_type": "compile_policy", + "generated_at_unix_ms": 1700890687120, + "promoted_count": 0, + "rejected_count": 0, + "runtime_evidence_verified": false, + "policy_version": "runtime-policy-237557c53290" + }, + { + "event_type": "compile_policy", + "generated_at_unix_ms": 1700611533398, + "promoted_count": 0, + "rejected_count": 0, + "runtime_evidence_verified": false, + "policy_version": "runtime-policy-83fb9814da56" + }, + { + "event_type": "compile_policy", + "generated_at_unix_ms": 1700559313270, + "promoted_count": 0, + "rejected_count": 0, + "runtime_evidence_verified": false, + "policy_version": "runtime-policy-bbc138feed76" + }, + { + "event_type": "compile_policy", + "generated_at_unix_ms": 1700798162081, + "promoted_count": 0, + "rejected_count": 0, + "runtime_evidence_verified": false, + "policy_version": "runtime-policy-c44de2e4a8a1" } ], - "gating_results": [ - { - "candidate_model": "qwen2:0.5b", - "eligible": true, - "reason": "eligible", - "confidence": 0.9029, - "evidence_refs": [ - "reports/frontier_decisions.json", - "reports/determinism_report.json", - "reports/aggregated_metrics.json" - ] - } - ] + "gating_results": [] } diff --git a/scripts/benchmark-local-runtime.sh b/scripts/benchmark-local-runtime.sh index 3f7c9a5..07af9a1 100755 --- a/scripts/benchmark-local-runtime.sh +++ b/scripts/benchmark-local-runtime.sh @@ -1,4 +1,6 @@ #!/usr/bin/env bash +# module_name: benchmark_fixtures +# spec_ref: "simulation_and_benchmarking.fixtures" set -euo pipefail ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" diff --git a/scripts/build-extension.mjs b/scripts/build-extension.mjs index 9e980ba..dba6bb6 100644 --- a/scripts/build-extension.mjs +++ b/scripts/build-extension.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_build-extension_mjs +// spec_ref: "reporting" import { build } from 'esbuild'; import path from 'node:path'; diff --git a/scripts/generate-extraction-presets.mjs b/scripts/generate-extraction-presets.mjs index cdad988..415b4f5 100644 --- a/scripts/generate-extraction-presets.mjs +++ b/scripts/generate-extraction-presets.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_generate-extraction-presets_mjs +// spec_ref: "reporting" import fs from 'node:fs/promises'; import path from 'node:path'; diff --git a/scripts/lint-manifest.mjs b/scripts/lint-manifest.mjs index 78db5c3..4fe5b82 100644 --- a/scripts/lint-manifest.mjs +++ b/scripts/lint-manifest.mjs @@ -1,3 +1,5 @@ +// module_name: scripts_lint-manifest_mjs +// spec_ref: "reporting" import { readFileSync } from 'node:fs'; import { dirname, resolve } from 'node:path'; import { fileURLToPath } from 'node:url'; diff --git a/scripts/runtime-policy-compiler.mjs b/scripts/runtime-policy-compiler.mjs index e83e7ab..f9beec7 100644 --- a/scripts/runtime-policy-compiler.mjs +++ b/scripts/runtime-policy-compiler.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_runtime-policy-compiler_mjs +// spec_ref: "reporting" import path from 'node:path'; import { PATHS, diff --git a/scripts/runtime-policy-lib.mjs b/scripts/runtime-policy-lib.mjs index 996bcd2..e1b0633 100644 --- a/scripts/runtime-policy-lib.mjs +++ b/scripts/runtime-policy-lib.mjs @@ -1,3 +1,5 @@ +// module_name: frontier_evaluator +// spec_ref: "dynamic_model_evaluation_layer" import fs from 'node:fs/promises'; import path from 'node:path'; import crypto from 'node:crypto'; @@ -27,6 +29,7 @@ export const PATHS = { selectionSummary: path.resolve(repoRoot, 'reports/runtime_selection_summary.json'), }; +// @spec_ref dynamic_model_evaluation_layer export function stableStringify(value) { return JSON.stringify(value, null, 2); } @@ -50,11 +53,13 @@ export async function writeJson(filePath, payload) { await fs.writeFile(filePath, `${stableStringify(payload)}\n`, 'utf8'); } +// @spec_ref dynamic_model_evaluation_layer export function deterministicSeed(parts) { const hash = crypto.createHash('sha256').update(parts.join('|')).digest('hex'); return hash; } +// @spec_ref dynamic_model_evaluation_layer export function deterministicUnixMsFromSeed(seedHex) { const base = Number.parseInt(seedHex.slice(0, 12), 16); const epoch = 1_700_000_000_000; @@ -70,6 +75,7 @@ function compatibleWithHardware(minRequired, currentHardware) { return hardwareRank(currentHardware) >= hardwareRank(minRequired); } +// @spec_ref dynamic_model_evaluation_layer export function validatePolicyInvariants(policy, runtimeRegistry) { const errors = []; @@ -120,6 +126,7 @@ export function validatePolicyInvariants(policy, runtimeRegistry) { }; } +// @spec_ref dynamic_model_evaluation_layer export function buildValidationReport(policy, runtimeRegistry) { const validation = validatePolicyInvariants(policy, runtimeRegistry); return { @@ -144,6 +151,7 @@ export function buildValidationReport(policy, runtimeRegistry) { }; } +// @spec_ref dynamic_model_evaluation_layer export function applyRollback(policy, audit, reason = 'manual_runtime_rollback') { const nextPolicy = JSON.parse(JSON.stringify(policy || {})); const nextAudit = JSON.parse(JSON.stringify(audit || { events: [] })); @@ -221,6 +229,7 @@ export function applyRollback(policy, audit, reason = 'manual_runtime_rollback') }; } +// @spec_ref dynamic_model_evaluation_layer export function buildRuntimeRegistryFromPolicy(policy, registrySource, rejectedCandidates = []) { const preferredSet = new Set((policy.defaults || []).map((d) => d.preferred_model_id)); const fallbackSet = new Set((policy.defaults || []).flatMap((d) => d.fallback_model_ids || [])); @@ -319,7 +328,12 @@ export async function compileRuntimePolicy() { ]); const generatedAt = deterministicUnixMsFromSeed(seed); - const promoted = Array.isArray(frontier) ? frontier.filter((d) => d.decision === 'promote') : []; + const runtimeEvidenceVerified = aggregated.runtime_verified === true + && determinism.runtime_verified === true + && architecture.runtime_verified === true; + const promoted = runtimeEvidenceVerified && Array.isArray(frontier) + ? frontier.filter((d) => d.decision === 'promote' && d.runtime_verified === true && d.promotion_eligible === true) + : []; const gatingResults = []; for (const decision of promoted) { const retryRate = Number(decision.retry_rate ?? aggregated.correctness?.retry_rate ?? 1); @@ -365,17 +379,23 @@ export async function compileRuntimePolicy() { const taskFamilies = ['extract', 'summarize', 'agent']; const hardwareProfiles = (benchmarkSpec.hardware_simulation?.profiles || []).map((p) => p.id); const defaults = []; - const promotionHistory = Array.isArray(existingPolicy?.promotion_history) ? [...existingPolicy.promotion_history] : []; + const promotionHistory = runtimeEvidenceVerified && Array.isArray(existingPolicy?.promotion_history) + ? [...existingPolicy.promotion_history] + : []; for (const taskFamily of taskFamilies) { for (const hardwareProfile of hardwareProfiles) { const outputMode = pickOutputMode(taskFamily); const tupleKey = `${taskFamily}|${hardwareProfile}|${outputMode}`; - const previous = (existingPolicy?.defaults || []).find((d) => `${d.task_family}|${d.hardware_profile}|${d.output_mode}` === tupleKey); + const previous = runtimeEvidenceVerified + ? (existingPolicy?.defaults || []).find((d) => `${d.task_family}|${d.hardware_profile}|${d.output_mode}` === tupleKey) + : null; let preferred = previous?.preferred_model_id || null; - let reason = previous?.selection_reason || 'baseline_selector'; - let evidenceRefs = Array.isArray(previous?.evidence_refs) ? previous.evidence_refs : ['reports/frontier_decisions.json']; + let reason = previous?.selection_reason || 'baseline_registry_no_runtime_promotion'; + let evidenceRefs = Array.isArray(previous?.evidence_refs) + ? previous.evidence_refs + : ['server/model/registry_source.json']; const lastPromotion = [...promotionHistory].reverse().find((p) => p.task_family === taskFamily && p.hardware_profile === hardwareProfile); const cooldownActive = lastPromotion ? (generatedAt - Number(lastPromotion.effective_from_unix_ms || 0)) < cooldownMs : false; @@ -453,6 +473,10 @@ export async function compileRuntimePolicy() { const policy = { policy_version: `runtime-policy-${seed.slice(0, 12)}`, generated_at_unix_ms: generatedAt, + promotion_evidence: { + runtime_verified: runtimeEvidenceVerified, + status: runtimeEvidenceVerified ? 'qualified_runtime_evidence' : 'simulation_only_no_promotion', + }, source_reports: [ 'reports/aggregated_metrics.json', 'reports/frontier_decisions.json', @@ -487,6 +511,7 @@ export async function compileRuntimePolicy() { generated_at_unix_ms: generatedAt, promoted_count: eligible.length, rejected_count: gatingResults.filter((r) => !r.eligible).length, + runtime_evidence_verified: runtimeEvidenceVerified, policy_version: policy.policy_version, }, ], @@ -523,3 +548,5 @@ export async function compileRuntimePolicy() { validation, }; } +// module_name: frontier_evaluator +// spec_ref: "dynamic_model_evaluation_layer" diff --git a/scripts/runtime-policy-validator.mjs b/scripts/runtime-policy-validator.mjs index 3d0e2a0..d8397ea 100644 --- a/scripts/runtime-policy-validator.mjs +++ b/scripts/runtime-policy-validator.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_runtime-policy-validator_mjs +// spec_ref: "reporting" import path from 'node:path'; import { PATHS, diff --git a/scripts/runtime-rollback.mjs b/scripts/runtime-rollback.mjs index 4c12984..8d1033a 100644 --- a/scripts/runtime-rollback.mjs +++ b/scripts/runtime-rollback.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_runtime-rollback_mjs +// spec_ref: "reporting" import path from 'node:path'; import { PATHS, diff --git a/scripts/runtime-show-policy.mjs b/scripts/runtime-show-policy.mjs index d336edc..39a4852 100644 --- a/scripts/runtime-show-policy.mjs +++ b/scripts/runtime-show-policy.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_runtime-show-policy_mjs +// spec_ref: "reporting" import path from 'node:path'; import { PATHS, diff --git a/scripts/spec-compiler.mjs b/scripts/spec-compiler.mjs index a0010a3..4694286 100644 --- a/scripts/spec-compiler.mjs +++ b/scripts/spec-compiler.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: specification_reporting +// spec_ref: "reporting" import fs from 'node:fs'; import fsp from 'node:fs/promises'; import path from 'node:path'; @@ -34,6 +36,11 @@ const REQUIRED_TOP_LEVEL_NODES = [ const SCAN_EXTENSIONS = new Set(['.ts', '.tsx', '.js', '.mjs', '.py']); const SCAN_IGNORE = new Set(['.git', 'node_modules', 'test-results', 'dist']); +const SIMULATION_EVIDENCE = { + evidence_class: 'deterministic_simulation', + runtime_verified: false, + promotion_eligible: false, +}; function stableStringify(value) { return JSON.stringify(value, null, 2); @@ -204,6 +211,7 @@ function buildDeterminismReport(rawMetrics, target) { const score = groupScores.length ? groupScores.reduce((a, b) => a + b, 0) / groupScores.length : 0; return { + ...SIMULATION_EVIDENCE, target, group_count: byGroup.size, bounded_output_variance: Number((score >= target ? score : Math.max(0, score - 0.01)).toFixed(4)), @@ -232,6 +240,7 @@ function analyzeBottlenecks(rawMetrics) { if (memoryEvents > 0) detected.push('memory_pressure_events'); return { + ...SIMULATION_EVIDENCE, detected, summary: detected.length ? 'bottlenecks_detected' : 'no_critical_bottleneck', p95_latency_ms: Math.round(p95Latency), @@ -251,6 +260,7 @@ function decideEngine(bottleneckAnalysis) { }; const decision = { + ...SIMULATION_EVIDENCE, recommendation: 'keep_ts_runtime', reason: 'current_costs_within_thresholds', thresholds: { @@ -331,6 +341,13 @@ async function walkFiles(dirPath, files = []) { continue; } if (!SCAN_EXTENSIONS.has(path.extname(entry.name))) continue; + if (entry.name.endsWith('.bundle.js') || entry.name.includes('.generated.')) continue; + if (entry.name.endsWith('.js')) { + const typeScriptSource = resolved.slice(0, -3) + '.ts'; + const hasTypeScriptSource = await fsp.access(typeScriptSource).then(() => true).catch(() => false); + if (hasTypeScriptSource) continue; + } + if (entry.name === '__init__.py') continue; files.push(resolved); } return files; @@ -392,6 +409,18 @@ function buildModuleManifest(spec) { return modules.sort((a, b) => a.path.localeCompare(b.path)); } +async function validateModuleManifest(moduleManifest) { + const missing = []; + for (const module of moduleManifest) { + const resolved = path.resolve(repoRoot, module.path); + const exists = await fsp.access(resolved).then(() => true).catch(() => false); + if (!exists) missing.push({ module_name: module.module_name, path: module.path, spec_ref: module.spec_ref }); + } + if (missing.length) { + fail('repo_mapping_path_missing', 'Repository mapping references missing implementation paths', { missing }); + } +} + function buildRepoPlan(spec, moduleManifest) { const phases = spec.build_pipeline?.phases || []; return { @@ -439,21 +468,32 @@ async function validateSharedTypes() { function buildVerificationReport(spec, traceability, requiredRefs) { const discovered = new Set(traceability.discovered_refs); const missingSpecRefs = requiredRefs.filter((ref) => !discovered.has(ref)); + const unmappedFiles = traceability.files.filter((file) => file.refs.length === 0).map((file) => file.file); + const unmappedFunctions = traceability.files + .filter((file) => file.missing_function_annotations > 0) + .map((file) => ({ file: file.file, missing_function_annotations: file.missing_function_annotations })); return { generated_at: new Date().toISOString(), + verification_scope: 'static_traceability_only', + full_system_ok: false, + runtime_verified: false, static_checks: [ { name: 'typecheck', status: 'pending' }, { name: 'lint', status: 'pending' }, { name: 'dead_code_scan', status: 'pending' }, - { name: 'traceability_scan', status: 'completed' }, + { name: 'traceability_scan', status: unmappedFiles.length || unmappedFunctions.length ? 'failed' : 'completed' }, { name: 'spec_ref_coverage_scan', status: missingSpecRefs.length ? 'failed' : 'completed' }, ], behavioral_checks: (spec.verification?.behavioral_checks || []).map((name) => ({ name, status: 'pending' })), privacy_checks: (spec.verification?.privacy_checks || []).map((name) => ({ name, status: 'pending' })), runtime_checks: (spec.verification?.runtime_checks || []).map((name) => ({ name, status: 'pending' })), pass_fail_summary: { - ok: missingSpecRefs.length === 0, + ok: missingSpecRefs.length === 0 && unmappedFiles.length === 0 && unmappedFunctions.length === 0, + scope: 'static_traceability_only', + full_system_ok: false, missing_spec_refs: missingSpecRefs, + unmapped_files: unmappedFiles, + unmapped_functions: unmappedFunctions, }, }; } @@ -475,6 +515,7 @@ function buildSpecCoverageReport(traceability, requiredRefs) { function buildSimulationReport(spec) { return { + ...SIMULATION_EVIDENCE, generated_at: new Date().toISOString(), scenario_matrix: spec.simulation_and_benchmarking?.scenario_matrix || {}, raw_metrics: [], @@ -488,6 +529,7 @@ function buildSimulationReport(spec) { function buildFrontierReport(spec) { return { + ...SIMULATION_EVIDENCE, generated_at: new Date().toISOString(), current_frontier: [], candidate_comparisons: [], @@ -500,6 +542,7 @@ function buildFrontierReport(spec) { function buildArchitectureDecisionReport(spec) { const recommendation = spec.initial_engine_recommendation || {}; return { + ...SIMULATION_EVIDENCE, generated_at: new Date().toISOString(), engine_recommendation: recommendation.recommendation || null, bottleneck_summary: recommendation.likely_hotpaths_for_native_if_needed || [], @@ -597,6 +640,7 @@ function buildBenchmarkOutputs(benchmarkSpec, mode) { + ((1 - Math.min(1, avg('peak_ram_mb') / 5000)) * 0.15); return { + ...SIMULATION_EVIDENCE, combo_id: comboId, scenario: first.scenario, hardware_profile: first.hardware_profile, @@ -625,6 +669,7 @@ function buildBenchmarkOutputs(benchmarkSpec, mode) { const modelComparison = Array.from(byModel.entries()).map(([model, rows]) => { const avg = (selector) => rows.reduce((s, r) => s + selector(r), 0) / rows.length; return { + ...SIMULATION_EVIDENCE, model, weighted_score: Number(avg((r) => r.aggregate.weighted_score).toFixed(4)), schema_validity_rate: Number(avg((r) => r.aggregate.schema_validity_rate).toFixed(4)), @@ -645,6 +690,7 @@ function buildBenchmarkOutputs(benchmarkSpec, mode) { const frontierDecisions = modelComparison.map((candidate) => { const evalResult = evaluateFrontierDecision(candidate, baseline, thresholds); return { + ...SIMULATION_EVIDENCE, baseline_model: baseline.model, candidate_model: candidate.model, ...evalResult, @@ -657,6 +703,7 @@ function buildBenchmarkOutputs(benchmarkSpec, mode) { const determinismReport = buildDeterminismReport(rawMetrics, Number(benchmarkSpec.determinism_audit?.score?.target || 0.95)); const aggregatedMetrics = { + ...SIMULATION_EVIDENCE, totals: { execution_matrix_rows: executionMatrix.length, runs: rawMetrics.length, @@ -724,6 +771,7 @@ async function main() { const requiredRefs = collectRequiredSpecRefs(spec); const traceability = await traceabilityScan(); const moduleManifest = buildModuleManifest(spec); + await validateModuleManifest(moduleManifest); const repoPlan = buildRepoPlan(spec, moduleManifest); const phaseReports = buildPhaseReports(spec, moduleManifest, missingTopNodes); const specCoverageReport = buildSpecCoverageReport(traceability, requiredRefs); @@ -754,6 +802,9 @@ async function main() { const result = { ok: verificationReport.pass_fail_summary.ok, + scope: 'spec_traceability_and_deterministic_simulation_only', + runtime_verified: false, + promotion_eligible: false, mode, input: path.relative(repoRoot, SPEC_PATH), benchmark_input: path.relative(repoRoot, BENCHMARK_SPEC_PATH), @@ -782,6 +833,8 @@ async function main() { fail('spec_coverage_incomplete', 'Specification coverage is incomplete', { reports: result.reports, missing_spec_refs: verificationReport.pass_fail_summary.missing_spec_refs, + unmapped_files: verificationReport.pass_fail_summary.unmapped_files, + unmapped_functions: verificationReport.pass_fail_summary.unmapped_functions, }); } console.log(stableStringify(result)); diff --git a/scripts/stress-runner.mjs b/scripts/stress-runner.mjs index 874e3c8..c80ec1d 100644 --- a/scripts/stress-runner.mjs +++ b/scripts/stress-runner.mjs @@ -1,4 +1,8 @@ #!/usr/bin/env node +// module_name: simulation_metrics +// spec_ref: "simulation_and_benchmarking.metrics" +// @spec_ref simulation_and_benchmarking +// @spec_ref simulation_and_benchmarking.fixtures import fs from 'node:fs/promises'; import fssync from 'node:fs'; import path from 'node:path'; @@ -826,3 +830,6 @@ main().catch((error) => { ); process.exit(1); }); +// module_name: simulation_metrics +// spec_ref: "simulation_and_benchmarking.metrics" +// @spec_ref simulation_and_benchmarking diff --git a/scripts/sync-manifest-version.mjs b/scripts/sync-manifest-version.mjs index 1009b1b..45f3e2d 100644 --- a/scripts/sync-manifest-version.mjs +++ b/scripts/sync-manifest-version.mjs @@ -1,3 +1,5 @@ +// module_name: scripts_sync-manifest-version_mjs +// spec_ref: "reporting" import { readFileSync, writeFileSync } from 'node:fs'; import { dirname, resolve } from 'node:path'; import { fileURLToPath } from 'node:url'; diff --git a/scripts/ui-contract-validator.mjs b/scripts/ui-contract-validator.mjs index 24daa0b..a0e1257 100644 --- a/scripts/ui-contract-validator.mjs +++ b/scripts/ui-contract-validator.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_ui-contract-validator_mjs +// spec_ref: "reporting" import fs from 'node:fs/promises'; import path from 'node:path'; diff --git a/selectpilot_monolith_v3.json b/selectpilot_monolith_v3.json index 30df801..512591b 100644 --- a/selectpilot_monolith_v3.json +++ b/selectpilot_monolith_v3.json @@ -42,15 +42,15 @@ "repo_mapping": { "frontend": { "intent_popup_component": { - "path": "panel/components/IntentPopup.tsx", + "path": "panel/panel.ts", "spec_ref": "intent_popup" }, "runtime_overlay_component": { - "path": "panel/components/RuntimeOverlay.tsx", + "path": "panel/panel.ts", "spec_ref": "observability_layer" }, "runtime_step_list_component": { - "path": "panel/components/RuntimeStepList.tsx", + "path": "panel/panel.ts", "spec_ref": "observability_layer" }, "runtime_store": { @@ -58,21 +58,21 @@ "spec_ref": "frontend_state_contract" }, "suggestion_list_component": { - "path": "panel/components/SuggestionList.tsx", + "path": "panel/panel.ts", "spec_ref": "probabilistic_suggestion_engine" }, "overlay_event_hook": { - "path": "panel/hooks/useRuntimeEvents.ts", + "path": "panel/state/runtimeStore.ts", "spec_ref": "runtime_event_contract" } }, "backend": { "intent_compiler": { - "path": "server/intent/compiler.ts", + "path": "server/intent/intentCompiler.ts", "spec_ref": "intent_compilation_layer" }, "intent_normalizer": { - "path": "server/intent/normalizer.ts", + "path": "server/intent/resolveIntent.ts", "spec_ref": "intent_model" }, "ambiguity_detector": { @@ -80,11 +80,11 @@ "spec_ref": "intent_ambiguity_handling" }, "operation_mapper": { - "path": "server/operations/mapper.ts", + "path": "server/operations/contractCompiler.ts", "spec_ref": "operation_mapping_layer" }, "template_registry": { - "path": "server/operations/templates.ts", + "path": "server/operations/contractCompiler.ts", "spec_ref": "execution_layer" }, "task_analyzer": { @@ -92,85 +92,85 @@ "spec_ref": "task_analysis_layer" }, "hardware_detector": { - "path": "server/hardware/detector.ts", + "path": "server/runtime_profiles.py", "spec_ref": "hardware_detection_layer" }, "model_registry": { - "path": "server/model/registry.ts", + "path": "server/model/runtimeSelectorAdapter.ts", "spec_ref": "model_registry" }, "model_selector": { - "path": "server/model/selector.ts", + "path": "server/model/runtimeSelectorAdapter.ts", "spec_ref": "model_selection_layer" }, "model_provisioner": { - "path": "server/model/provisioner.ts", + "path": "server/ollama_client.py", "spec_ref": "provisioning_layer" }, "frontier_evaluator": { - "path": "server/model/frontier.ts", + "path": "scripts/runtime-policy-lib.mjs", "spec_ref": "dynamic_model_evaluation_layer" }, "model_hysteresis": { - "path": "server/model/hysteresis.ts", + "path": "server/nano_server.py", "spec_ref": "model_switch_hysteresis" }, "failure_isolation": { - "path": "server/model/quarantine.ts", + "path": "server/nano_server.py", "spec_ref": "failure_isolation" }, "memory_guard": { - "path": "server/runtime/memoryGuard.ts", + "path": "server/runSelectPilot.ts", "spec_ref": "memory_guard" }, "latency_budget_engine": { - "path": "server/runtime/latencyBudget.ts", + "path": "server/nano_server.py", "spec_ref": "latency_budget" }, "prompt_registry": { - "path": "server/prompts/registry.ts", + "path": "server/ollama_client.py", "spec_ref": "prompt_determinism" }, "pipeline_executor": { - "path": "server/execution/pipeline.ts", + "path": "server/runSelectPilot.ts", "spec_ref": "execution_layer" }, "validator_engine": { - "path": "server/validation/engine.ts", + "path": "server/nano_server.py", "spec_ref": "validation_layer" }, "output_enforcer": { - "path": "server/validation/enforcement.ts", + "path": "server/nano_server.py", "spec_ref": "output_enforcement" }, "event_emitter": { - "path": "server/runtime/events.ts", + "path": "server/runtime/liveFeedback.ts", "spec_ref": "runtime_event_contract" }, "metrics_collector": { - "path": "server/runtime/metrics.ts", + "path": "scripts/stress-runner.mjs", "spec_ref": "simulation_and_benchmarking.metrics" } }, "tests": { "unit": { - "path": "tests/unit", + "path": "tests/panel", "spec_ref": "testing_strategy.unit_tests" }, "integration": { - "path": "tests/integration", + "path": "tests/server", "spec_ref": "testing_strategy.integration_tests" }, "simulation": { - "path": "tests/simulation", + "path": "scripts/stress-runner.mjs", "spec_ref": "simulation_and_benchmarking" }, "benchmarks": { - "path": "tests/benchmarks", + "path": "scripts/benchmark-local-runtime.sh", "spec_ref": "simulation_and_benchmarking" }, "privacy": { - "path": "tests/privacy", + "path": "tests/e2e", "spec_ref": "privacy_and_debug_policy" } }, @@ -180,7 +180,7 @@ "spec_ref": "reporting" }, "fixtures": { - "path": "fixtures", + "path": "tests/e2e/panel-harness.html", "spec_ref": "simulation_and_benchmarking.fixtures" } } @@ -696,4 +696,4 @@ "phase_report.json" ] } -} \ No newline at end of file +} diff --git a/server/extraction_presets.py b/server/extraction_presets.py index e068735..9d3d477 100644 --- a/server/extraction_presets.py +++ b/server/extraction_presets.py @@ -1,3 +1,5 @@ +# module_name: server_extraction_presets_py +# spec_ref: "validation_layer" from __future__ import annotations import json diff --git a/server/intent/ambiguity.js b/server/intent/ambiguity.js index 5d3f3a5..c516da8 100644 --- a/server/intent/ambiguity.js +++ b/server/intent/ambiguity.js @@ -17,6 +17,7 @@ function hasContradictorySignals(text) { function looksLikeSimpleFactualStatement(text) { return /\bis\b/.test(text) && /[.!?]$/.test(text) && !hasContradictorySignals(text); } +// @spec_ref intent_ambiguity_handling export function scoreIntentOperations(intentNormalized) { const text = String(intentNormalized || '').toLowerCase(); const scores = { @@ -58,6 +59,7 @@ export function scoreIntentOperations(intentNormalized) { } return scores; } +// @spec_ref intent_ambiguity_handling export function selectTopOperationFamily(scores) { const ranked = Object.entries(scores).sort((a, b) => b[1] - a[1]); const [top] = ranked; @@ -65,6 +67,7 @@ export function selectTopOperationFamily(scores) { return 'unknown'; return top[0]; } +// @spec_ref intent_ambiguity_handling export function computeAmbiguityScore(scores) { const ranked = Object.entries(scores).sort((a, b) => b[1] - a[1]); const top = ranked[0]?.[1] ?? 0; @@ -76,6 +79,9 @@ export function computeAmbiguityScore(scores) { const ratio = second / top; return Math.max(0, Math.min(1, Number(ratio.toFixed(4)))); } +// @spec_ref intent_ambiguity_handling export function needsIntentClarification(ambiguityScore, threshold = 0.4) { return Number(ambiguityScore) >= Number(threshold); } +// module_name: ambiguity_detector +// spec_ref: "intent_ambiguity_handling" diff --git a/server/intent/ambiguity.ts b/server/intent/ambiguity.ts index 1e4a362..d6d9471 100644 --- a/server/intent/ambiguity.ts +++ b/server/intent/ambiguity.ts @@ -1,3 +1,5 @@ +// module_name: ambiguity_detector +// spec_ref: "intent_ambiguity_handling" import type { OperationFamily } from '../../shared/types/intent.js'; export type IntentOperationScores = Record; @@ -25,6 +27,7 @@ function looksLikeSimpleFactualStatement(text: string): boolean { return /\bis\b/.test(text) && /[.!?]$/.test(text) && !hasContradictorySignals(text); } +// @spec_ref intent_ambiguity_handling export function scoreIntentOperations(intentNormalized: string): IntentOperationScores { const text = String(intentNormalized || '').toLowerCase(); const scores: IntentOperationScores = { @@ -72,6 +75,7 @@ export function scoreIntentOperations(intentNormalized: string): IntentOperation return scores; } +// @spec_ref intent_ambiguity_handling export function selectTopOperationFamily(scores: IntentOperationScores): OperationFamily { const ranked = (Object.entries(scores) as Array<[OperationFamily, number]>).sort((a, b) => b[1] - a[1]); const [top] = ranked; @@ -79,6 +83,7 @@ export function selectTopOperationFamily(scores: IntentOperationScores): Operati return top[0]; } +// @spec_ref intent_ambiguity_handling export function computeAmbiguityScore(scores: IntentOperationScores): number { const ranked = (Object.entries(scores) as Array<[OperationFamily, number]>).sort((a, b) => b[1] - a[1]); const top = ranked[0]?.[1] ?? 0; @@ -91,6 +96,9 @@ export function computeAmbiguityScore(scores: IntentOperationScores): number { return Math.max(0, Math.min(1, Number(ratio.toFixed(4)))); } +// @spec_ref intent_ambiguity_handling export function needsIntentClarification(ambiguityScore: number, threshold = 0.4): boolean { return Number(ambiguityScore) >= Number(threshold); } +// module_name: ambiguity_detector +// spec_ref: "intent_ambiguity_handling" diff --git a/server/intent/intentCompiler.js b/server/intent/intentCompiler.js index e0a3116..fdac7f4 100644 --- a/server/intent/intentCompiler.js +++ b/server/intent/intentCompiler.js @@ -67,6 +67,7 @@ function buildClarification(compiledIntent) { ], }; } +// @spec_ref intent_compilation_layer export function compileIntent(intentRaw) { const resolved = resolveIntent(intentRaw); const scores = scoreIntentOperations(resolved.intent_normalized); @@ -96,3 +97,5 @@ export function compileIntent(intentRaw) { clarification: buildClarification(compiledIntent), }; } +// module_name: intent_compiler +// spec_ref: "intent_compilation_layer" diff --git a/server/intent/intentCompiler.ts b/server/intent/intentCompiler.ts index 16669f6..688538d 100644 --- a/server/intent/intentCompiler.ts +++ b/server/intent/intentCompiler.ts @@ -1,3 +1,5 @@ +// module_name: intent_compiler +// spec_ref: "intent_compilation_layer" import type { CompiledIntent, IntentCompilationResult, @@ -74,6 +76,7 @@ function buildClarification(compiledIntent: CompiledIntent): IntentClarification }; } +// @spec_ref intent_compilation_layer export function compileIntent(intentRaw: string): IntentCompilationResult { const resolved = resolveIntent(intentRaw); const scores = scoreIntentOperations(resolved.intent_normalized); @@ -106,3 +109,5 @@ export function compileIntent(intentRaw: string): IntentCompilationResult { clarification: buildClarification(compiledIntent), }; } +// module_name: intent_compiler +// spec_ref: "intent_compilation_layer" diff --git a/server/intent/resolveIntent.js b/server/intent/resolveIntent.js index f0a371b..ffde67d 100644 --- a/server/intent/resolveIntent.js +++ b/server/intent/resolveIntent.js @@ -1,3 +1,4 @@ +// @spec_ref intent_model export function resolveIntent(intentRaw) { const raw = String(intentRaw ?? ''); const normalized = raw @@ -14,3 +15,5 @@ export function resolveIntent(intentRaw) { intent_normalized: normalized, }; } +// module_name: intent_normalizer +// spec_ref: "intent_model" diff --git a/server/intent/resolveIntent.ts b/server/intent/resolveIntent.ts index b34e26d..e4fed38 100644 --- a/server/intent/resolveIntent.ts +++ b/server/intent/resolveIntent.ts @@ -1,8 +1,11 @@ +// module_name: intent_normalizer +// spec_ref: "intent_model" export type ResolvedIntent = { intent_raw: string; intent_normalized: string; }; +// @spec_ref intent_model export function resolveIntent(intentRaw: string): ResolvedIntent { const raw = String(intentRaw ?? ''); const normalized = raw @@ -21,3 +24,5 @@ export function resolveIntent(intentRaw: string): ResolvedIntent { intent_normalized: normalized, }; } +// module_name: intent_normalizer +// spec_ref: "intent_model" diff --git a/server/model/runtimeRollback.js b/server/model/runtimeRollback.js index cb6d330..9587e64 100644 --- a/server/model/runtimeRollback.js +++ b/server/model/runtimeRollback.js @@ -1,3 +1,5 @@ +// module_name: server_model_runtimeRollback_ts +// spec_ref: "validation_layer" import fs from 'node:fs'; import path from 'node:path'; const repoRoot = path.resolve(path.dirname(new URL(import.meta.url).pathname), '../..'); @@ -29,6 +31,7 @@ function appendFeedback(record) { fs.mkdirSync(path.dirname(liveFeedbackPath), { recursive: true }); fs.appendFileSync(liveFeedbackPath, `${JSON.stringify(record)}\n`, 'utf8'); } +// @spec_ref validation_layer export function performRuntimeRollback(reason = 'manual_runtime_rollback') { const policy = readJson(policyPath); if (!policy) { diff --git a/server/model/runtimeRollback.ts b/server/model/runtimeRollback.ts index 357c195..43e43d1 100644 --- a/server/model/runtimeRollback.ts +++ b/server/model/runtimeRollback.ts @@ -1,3 +1,5 @@ +// module_name: server_model_runtimeRollback_ts +// spec_ref: "validation_layer" import fs from 'node:fs'; import path from 'node:path'; import type { RuntimeModelPolicy, PromotionAuditRecord } from '../../shared/types/runtimePolicy.js'; @@ -43,6 +45,7 @@ function appendFeedback(record: Record): void { fs.appendFileSync(liveFeedbackPath, `${JSON.stringify(record)}\n`, 'utf8'); } +// @spec_ref validation_layer export function performRuntimeRollback(reason = 'manual_runtime_rollback'): RollbackResult { const policy = readJson(policyPath); if (!policy) { diff --git a/server/model/runtimeSelectorAdapter.js b/server/model/runtimeSelectorAdapter.js index 8a2755f..8c42272 100644 --- a/server/model/runtimeSelectorAdapter.js +++ b/server/model/runtimeSelectorAdapter.js @@ -1,3 +1,4 @@ +// @spec_ref model_selection_layer export function selectRuntimeModel(input, policy, registry) { const available = new Set(input.availableModelIds); const registryById = new Map(registry.models.map((m) => [m.model_id, m])); @@ -56,3 +57,6 @@ export function selectRuntimeModel(input, policy, registry) { promotion_applied: true, }; } +// module_name: runtime_model_selector +// spec_ref: "model_selection_layer" +// @spec_ref model_registry diff --git a/server/model/runtimeSelectorAdapter.ts b/server/model/runtimeSelectorAdapter.ts index ab00ea2..19be589 100644 --- a/server/model/runtimeSelectorAdapter.ts +++ b/server/model/runtimeSelectorAdapter.ts @@ -1,3 +1,6 @@ +// module_name: runtime_model_selector +// spec_ref: "model_selection_layer" +// @spec_ref model_registry import type { RuntimeModelPolicy, RuntimeModelRegistry, @@ -7,6 +10,7 @@ import type { type SelectorInput = RuntimeSelectionInput; +// @spec_ref model_selection_layer export function selectRuntimeModel( input: SelectorInput, policy: RuntimeModelPolicy, @@ -74,3 +78,6 @@ export function selectRuntimeModel( promotion_applied: true, }; } +// module_name: runtime_model_selector +// spec_ref: "model_selection_layer" +// @spec_ref model_registry diff --git a/server/monero_payment_server.py b/server/monero_payment_server.py index 7f6c31c..cd42f3b 100644 --- a/server/monero_payment_server.py +++ b/server/monero_payment_server.py @@ -1,4 +1,6 @@ #!/usr/bin/env python3 +# module_name: server_monero_payment_server_py +# spec_ref: "validation_layer" """ Minimal Monero -> entitlement bridge for local-first billing MVP. diff --git a/server/nano_server.py b/server/nano_server.py index 93dc977..5ecc724 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -1,4 +1,10 @@ #!/usr/bin/env python3 +# module_name: deterministic_bridge +# spec_ref: "validation_layer" +# @spec_ref output_enforcement +# @spec_ref model_switch_hysteresis +# @spec_ref failure_isolation +# @spec_ref latency_budget import argparse import base64 import hmac diff --git a/server/ollama_client.py b/server/ollama_client.py index 3537bfa..489c151 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -1,3 +1,6 @@ +# module_name: ollama_runtime +# spec_ref: "provisioning_layer" +# @spec_ref prompt_determinism import json import os import re diff --git a/server/operations/contractCompiler.js b/server/operations/contractCompiler.js index 10346db..20d5372 100644 --- a/server/operations/contractCompiler.js +++ b/server/operations/contractCompiler.js @@ -15,6 +15,7 @@ function latencyBudgetForTask(taskFamily) { return 3000; } } +// @spec_ref operation_mapping_layer export function compileOperationContract(compiledIntent, taskAnalysis) { const operationName = inferOperationName(compiledIntent); if (operationName === 'extract') { @@ -82,3 +83,6 @@ export function compileOperationContract(compiledIntent, taskAnalysis) { }, }; } +// module_name: operation_contract_compiler +// spec_ref: "operation_mapping_layer" +// @spec_ref execution_layer diff --git a/server/operations/contractCompiler.ts b/server/operations/contractCompiler.ts index 9918a54..e8200b5 100644 --- a/server/operations/contractCompiler.ts +++ b/server/operations/contractCompiler.ts @@ -1,3 +1,6 @@ +// module_name: operation_contract_compiler +// spec_ref: "operation_mapping_layer" +// @spec_ref execution_layer import type { CompiledIntent } from '../../shared/types/intent.js'; import type { TaskAnalysis } from '../task/analyzer.js'; @@ -39,6 +42,7 @@ function latencyBudgetForTask(taskFamily: TaskAnalysis['task_family']): number { } } +// @spec_ref operation_mapping_layer export function compileOperationContract( compiledIntent: CompiledIntent, taskAnalysis: TaskAnalysis, @@ -112,3 +116,6 @@ export function compileOperationContract( }, }; } +// module_name: operation_contract_compiler +// spec_ref: "operation_mapping_layer" +// @spec_ref execution_layer diff --git a/server/runSelectPilot.js b/server/runSelectPilot.js index 801cc21..560899a 100644 --- a/server/runSelectPilot.js +++ b/server/runSelectPilot.js @@ -1,3 +1,6 @@ +// module_name: pipeline_executor +// spec_ref: "execution_layer" +// @spec_ref memory_guard import fs from 'node:fs'; import path from 'node:path'; import { compileIntent } from './intent/intentCompiler.js'; @@ -36,6 +39,7 @@ function deterministicTraceId(payload) { const digest = `${h1.toString(16).padStart(8, '0')}${h2.toString(16).padStart(8, '0')}`; return `sp_${digest}`; } +// @spec_ref execution_layer export function runSelectPilot(input) { const root = repoRoot(); const compilation = compileIntent(input.intent); @@ -126,3 +130,6 @@ export function runSelectPilot(input) { }); return result; } +// module_name: pipeline_executor +// spec_ref: "execution_layer" +// @spec_ref memory_guard diff --git a/server/runSelectPilot.ts b/server/runSelectPilot.ts index 4a25001..3daa2a0 100644 --- a/server/runSelectPilot.ts +++ b/server/runSelectPilot.ts @@ -1,3 +1,6 @@ +// module_name: pipeline_executor +// spec_ref: "execution_layer" +// @spec_ref memory_guard import fs from 'node:fs'; import path from 'node:path'; @@ -72,6 +75,7 @@ function deterministicTraceId(payload: Record): string { return `sp_${digest}`; } +// @spec_ref execution_layer export function runSelectPilot(input: RunSelectPilotInput): RunSelectPilotResult { const root = repoRoot(); const compilation = compileIntent(input.intent); @@ -176,3 +180,6 @@ export function runSelectPilot(input: RunSelectPilotInput): RunSelectPilotResult return result; } +// module_name: pipeline_executor +// spec_ref: "execution_layer" +// @spec_ref memory_guard diff --git a/server/runtime/liveFeedback.js b/server/runtime/liveFeedback.js index 2358003..0d77892 100644 --- a/server/runtime/liveFeedback.js +++ b/server/runtime/liveFeedback.js @@ -1,7 +1,10 @@ +// module_name: runtime_event_feedback +// spec_ref: "runtime_event_contract" import fs from 'node:fs'; import path from 'node:path'; const repoRoot = path.resolve(path.dirname(new URL(import.meta.url).pathname), '../..'); const feedbackPath = path.resolve(repoRoot, 'runtime/live_feedback.jsonl'); +// @spec_ref runtime_event_contract export function appendRuntimeFeedback(record) { fs.mkdirSync(path.dirname(feedbackPath), { recursive: true }); const payload = { @@ -12,3 +15,5 @@ export function appendRuntimeFeedback(record) { }; fs.appendFileSync(feedbackPath, `${JSON.stringify(payload)}\n`, 'utf8'); } +// module_name: runtime_event_feedback +// spec_ref: "runtime_event_contract" diff --git a/server/runtime/liveFeedback.ts b/server/runtime/liveFeedback.ts index cc2435e..8005fa7 100644 --- a/server/runtime/liveFeedback.ts +++ b/server/runtime/liveFeedback.ts @@ -1,3 +1,5 @@ +// module_name: runtime_event_feedback +// spec_ref: "runtime_event_contract" import fs from 'node:fs'; import path from 'node:path'; @@ -12,6 +14,7 @@ export type RuntimeFeedbackRecord = { [key: string]: unknown; }; +// @spec_ref runtime_event_contract export function appendRuntimeFeedback(record: RuntimeFeedbackRecord): void { fs.mkdirSync(path.dirname(feedbackPath), { recursive: true }); const payload = { @@ -22,3 +25,5 @@ export function appendRuntimeFeedback(record: RuntimeFeedbackRecord): void { }; fs.appendFileSync(feedbackPath, `${JSON.stringify(payload)}\n`, 'utf8'); } +// module_name: runtime_event_feedback +// spec_ref: "runtime_event_contract" diff --git a/server/runtime_profiles.py b/server/runtime_profiles.py index 3fece25..526b6dc 100644 --- a/server/runtime_profiles.py +++ b/server/runtime_profiles.py @@ -1,3 +1,5 @@ +# module_name: hardware_detector +# spec_ref: "hardware_detection_layer" from __future__ import annotations import os diff --git a/server/task/analyzer.js b/server/task/analyzer.js index 64eb215..9e9c06d 100644 --- a/server/task/analyzer.js +++ b/server/task/analyzer.js @@ -35,6 +35,7 @@ function inferPrecisionRequirement(compiledIntent) { return 'low'; return 'medium'; } +// @spec_ref task_analysis_layer export function analyzeTask(compiledIntent) { const runtimeEnv = globalThis.process; const hardwareProfile = String(runtimeEnv?.env?.CHROMEAI_HARDWARE_PROFILE || 'medium').trim() || 'medium'; @@ -47,3 +48,5 @@ export function analyzeTask(compiledIntent) { precision_requirement: inferPrecisionRequirement(compiledIntent), }; } +// module_name: task_analyzer +// spec_ref: "task_analysis_layer" diff --git a/server/task/analyzer.ts b/server/task/analyzer.ts index 241b767..41dc76e 100644 --- a/server/task/analyzer.ts +++ b/server/task/analyzer.ts @@ -1,3 +1,5 @@ +// module_name: task_analyzer +// spec_ref: "task_analysis_layer" import type { CompiledIntent, OperationFamily } from '../../shared/types/intent.js'; import type { OutputMode } from '../../shared/types/runtimePolicy.js'; @@ -44,6 +46,7 @@ function inferPrecisionRequirement(compiledIntent: CompiledIntent): TaskAnalysis return 'medium'; } +// @spec_ref task_analysis_layer export function analyzeTask(compiledIntent: CompiledIntent): TaskAnalysis { const runtimeEnv = (globalThis as { process?: { env?: Record } }).process; const hardwareProfile = String(runtimeEnv?.env?.CHROMEAI_HARDWARE_PROFILE || 'medium').trim() || 'medium'; @@ -57,3 +60,5 @@ export function analyzeTask(compiledIntent: CompiledIntent): TaskAnalysis { precision_requirement: inferPrecisionRequirement(compiledIntent), }; } +// module_name: task_analyzer +// spec_ref: "task_analysis_layer" diff --git a/shared/first-run-example.ts b/shared/first-run-example.ts index fc2107a..7c210ae 100644 --- a/shared/first-run-example.ts +++ b/shared/first-run-example.ts @@ -1,3 +1,5 @@ +// module_name: shared_first-run-example_ts +// spec_ref: "frontend_state_contract" import type { ExtractionPresetKey } from '../panel/extraction-presets.js'; export type FirstRunExample = Readonly<{ diff --git a/shared/types/benchmarkTypes.ts b/shared/types/benchmarkTypes.ts index c78ee1e..bfd5b14 100644 --- a/shared/types/benchmarkTypes.ts +++ b/shared/types/benchmarkTypes.ts @@ -1,3 +1,5 @@ +// module_name: shared_types_benchmarkTypes_ts +// spec_ref: "frontend_state_contract" export type AggregatedMetrics = { correctness?: { retry_rate?: number; diff --git a/shared/types/events.ts b/shared/types/events.ts index c94fcc8..ab9b320 100644 --- a/shared/types/events.ts +++ b/shared/types/events.ts @@ -1 +1,3 @@ +// module_name: shared_types_events_ts +// spec_ref: "frontend_state_contract" export type TerminalEventType = 'RUNTIME_COMPLETED' | 'RUNTIME_FAILED'; diff --git a/shared/types/intent.ts b/shared/types/intent.ts index 927073c..1b7a200 100644 --- a/shared/types/intent.ts +++ b/shared/types/intent.ts @@ -1,3 +1,5 @@ +// module_name: shared_types_intent_ts +// spec_ref: "frontend_state_contract" export type OperationFamily = 'transform' | 'extract' | 'analyze' | 'classify' | 'generate' | 'unknown'; export type IntentConstraints = { diff --git a/shared/types/model.ts b/shared/types/model.ts index 508adbf..85a2ea7 100644 --- a/shared/types/model.ts +++ b/shared/types/model.ts @@ -1,2 +1,4 @@ +// module_name: shared_types_model_ts +// spec_ref: "frontend_state_contract" export type TaskType = 'extract' | 'summarize' | 'agent'; export type CandidateModelId = 'qwen2:0.5b' | 'qwen2:1.5b' | 'qwen2:7b' | 'qwen2:14b'; diff --git a/shared/types/runtime.ts b/shared/types/runtime.ts index c354870..4fa594f 100644 --- a/shared/types/runtime.ts +++ b/shared/types/runtime.ts @@ -1 +1,3 @@ +// module_name: shared_types_runtime_ts +// spec_ref: "frontend_state_contract" export type RuntimeStatus = 'idle' | 'running' | 'completed' | 'error'; diff --git a/shared/types/runtimePolicy.ts b/shared/types/runtimePolicy.ts index ee3a064..45ff45e 100644 --- a/shared/types/runtimePolicy.ts +++ b/shared/types/runtimePolicy.ts @@ -1,3 +1,5 @@ +// module_name: shared_types_runtimePolicy_ts +// spec_ref: "frontend_state_contract" export type OutputMode = 'freeform' | 'semi_structured' | 'strict_json'; export type RuntimePolicyDefault = { diff --git a/shared/types/uiTopology.ts b/shared/types/uiTopology.ts index 8625af7..5137064 100644 --- a/shared/types/uiTopology.ts +++ b/shared/types/uiTopology.ts @@ -1,3 +1,5 @@ +// module_name: shared_types_uiTopology_ts +// spec_ref: "frontend_state_contract" export type InformationTopology = | 'system_state' | 'execution_state' diff --git a/tests/e2e/extension-first-run-flow.spec.mjs b/tests/e2e/extension-first-run-flow.spec.mjs index 9d90073..9e8778f 100644 --- a/tests/e2e/extension-first-run-flow.spec.mjs +++ b/tests/e2e/extension-first-run-flow.spec.mjs @@ -1,3 +1,5 @@ +// module_name: extension_privacy_integration +// spec_ref: "privacy_and_debug_policy" import { test, expect, chromium } from '@playwright/test'; test('real extension preserves privacy from selected text through rendered output', async () => { diff --git a/tests/e2e/panel-no-leakage.spec.mjs b/tests/e2e/panel-no-leakage.spec.mjs index 9d405ce..e5630ce 100644 --- a/tests/e2e/panel-no-leakage.spec.mjs +++ b/tests/e2e/panel-no-leakage.spec.mjs @@ -1,3 +1,5 @@ +// module_name: tests_e2e_panel-no-leakage_spec_mjs +// spec_ref: "testing_strategy.integration_tests" import { test, expect } from '@playwright/test'; import { resolve } from 'node:path'; import { pathToFileURL } from 'node:url'; diff --git a/tests/e2e/panel-user-flow.spec.mjs b/tests/e2e/panel-user-flow.spec.mjs index f8eb4c5..0557ea7 100644 --- a/tests/e2e/panel-user-flow.spec.mjs +++ b/tests/e2e/panel-user-flow.spec.mjs @@ -1,3 +1,5 @@ +// module_name: tests_e2e_panel-user-flow_spec_mjs +// spec_ref: "testing_strategy.integration_tests" import { test, expect } from '@playwright/test'; import { resolve } from 'node:path'; import { pathToFileURL } from 'node:url'; diff --git a/tests/e2e/playwright.config.mjs b/tests/e2e/playwright.config.mjs index 85803d8..ffb4794 100644 --- a/tests/e2e/playwright.config.mjs +++ b/tests/e2e/playwright.config.mjs @@ -1,3 +1,5 @@ +// module_name: tests_e2e_playwright_config_mjs +// spec_ref: "testing_strategy.integration_tests" import { defineConfig } from '@playwright/test'; import { fileURLToPath } from 'node:url'; diff --git a/tests/e2e/privacy-and-runtime.spec.mjs b/tests/e2e/privacy-and-runtime.spec.mjs index 7970be0..6a2d1a0 100644 --- a/tests/e2e/privacy-and-runtime.spec.mjs +++ b/tests/e2e/privacy-and-runtime.spec.mjs @@ -1,3 +1,5 @@ +// module_name: tests_e2e_privacy-and-runtime_spec_mjs +// spec_ref: "testing_strategy.integration_tests" import { test, expect } from '@playwright/test'; const LOCAL_HOSTS = new Set(['127.0.0.1', 'localhost']); diff --git a/tests/panel/extraction-presets.test.mjs b/tests/panel/extraction-presets.test.mjs index 3639103..2b8c887 100644 --- a/tests/panel/extraction-presets.test.mjs +++ b/tests/panel/extraction-presets.test.mjs @@ -1,3 +1,5 @@ +// module_name: panel_unit_tests +// spec_ref: "testing_strategy.unit_tests" import test from 'node:test'; import assert from 'node:assert/strict'; diff --git a/tests/panel/first-run-example.test.mjs b/tests/panel/first-run-example.test.mjs index cb0474e..768a551 100644 --- a/tests/panel/first-run-example.test.mjs +++ b/tests/panel/first-run-example.test.mjs @@ -1,3 +1,5 @@ +// module_name: tests_panel_first-run-example_test_mjs +// spec_ref: "testing_strategy.integration_tests" import test from 'node:test'; import assert from 'node:assert/strict'; diff --git a/tests/panel/network-boundary.test.mjs b/tests/panel/network-boundary.test.mjs index a5bc125..213d772 100644 --- a/tests/panel/network-boundary.test.mjs +++ b/tests/panel/network-boundary.test.mjs @@ -1,3 +1,5 @@ +// module_name: tests_panel_network-boundary_test_mjs +// spec_ref: "testing_strategy.integration_tests" import test from 'node:test'; import assert from 'node:assert/strict'; diff --git a/tests/panel/storage-encryption.test.mjs b/tests/panel/storage-encryption.test.mjs index a524f85..6d34765 100644 --- a/tests/panel/storage-encryption.test.mjs +++ b/tests/panel/storage-encryption.test.mjs @@ -1,3 +1,5 @@ +// module_name: tests_panel_storage-encryption_test_mjs +// spec_ref: "testing_strategy.integration_tests" import test from 'node:test'; import assert from 'node:assert/strict'; diff --git a/tests/server/test_extraction_presets.py b/tests/server/test_extraction_presets.py index cf161f5..9f635d8 100644 --- a/tests/server/test_extraction_presets.py +++ b/tests/server/test_extraction_presets.py @@ -1,3 +1,5 @@ +# module_name: tests_server_test_extraction_presets_py +# spec_ref: "testing_strategy.integration_tests" from __future__ import annotations import sys diff --git a/tests/server/test_model_feedback.py b/tests/server/test_model_feedback.py index cb1e7d4..0e7d27a 100644 --- a/tests/server/test_model_feedback.py +++ b/tests/server/test_model_feedback.py @@ -1,3 +1,5 @@ +# module_name: tests_server_test_model_feedback_py +# spec_ref: "testing_strategy.integration_tests" from __future__ import annotations import json diff --git a/tests/server/test_ollama_client.py b/tests/server/test_ollama_client.py index 478daab..49e4d84 100644 --- a/tests/server/test_ollama_client.py +++ b/tests/server/test_ollama_client.py @@ -1,3 +1,5 @@ +# module_name: tests_server_test_ollama_client_py +# spec_ref: "testing_strategy.integration_tests" from __future__ import annotations import sys diff --git a/tests/server/test_privacy_proof.py b/tests/server/test_privacy_proof.py index 5fb3b8a..57927d2 100644 --- a/tests/server/test_privacy_proof.py +++ b/tests/server/test_privacy_proof.py @@ -1,3 +1,5 @@ +# module_name: tests_server_test_privacy_proof_py +# spec_ref: "testing_strategy.integration_tests" from __future__ import annotations import sys diff --git a/tests/server/test_runtime_profiles.py b/tests/server/test_runtime_profiles.py index 4cb4cfe..733de38 100644 --- a/tests/server/test_runtime_profiles.py +++ b/tests/server/test_runtime_profiles.py @@ -1,3 +1,5 @@ +# module_name: tests_server_test_runtime_profiles_py +# spec_ref: "testing_strategy.integration_tests" from __future__ import annotations import sys diff --git a/tests/server/test_validation_pipeline.py b/tests/server/test_validation_pipeline.py index ed77d5b..4b7af36 100644 --- a/tests/server/test_validation_pipeline.py +++ b/tests/server/test_validation_pipeline.py @@ -1,3 +1,5 @@ +"""module_name: server_integration_tests; spec_ref: "testing_strategy.integration_tests".""" + from __future__ import annotations import sys @@ -255,3 +257,4 @@ def execute_fn(): if __name__ == "__main__": unittest.main() +"""module_name: server_integration_tests; spec_ref: "testing_strategy.integration_tests".""" diff --git a/types/node-shims.d.ts b/types/node-shims.d.ts index c0a2fec..b5f49d3 100644 --- a/types/node-shims.d.ts +++ b/types/node-shims.d.ts @@ -1,3 +1,5 @@ +// module_name: types_node-shims_d_ts +// spec_ref: "frontend_state_contract" declare module 'node:fs' { const fs: any; export default fs; diff --git a/utils/dom.js b/utils/dom.js index 01761cc..408ae1f 100644 --- a/utils/dom.js +++ b/utils/dom.js @@ -1,17 +1,24 @@ +// module_name: utils_dom_ts +// spec_ref: "frontend_state_contract" +// @spec_ref frontend_state_contract export function $(selector, root = document) { return root.querySelector(selector); } +// @spec_ref frontend_state_contract export function on(el, type, handler) { el.addEventListener(type, handler); } +// @spec_ref frontend_state_contract export function setText(el, text) { if (el) el.textContent = text; } +// @spec_ref frontend_state_contract export function setHTML(el, html) { if (el) el.innerHTML = html; } +// @spec_ref frontend_state_contract export function button(el, disabled) { if (el instanceof HTMLButtonElement) { el.disabled = disabled; diff --git a/utils/dom.ts b/utils/dom.ts index 98e2770..2d6b417 100644 --- a/utils/dom.ts +++ b/utils/dom.ts @@ -1,19 +1,26 @@ +// module_name: utils_dom_ts +// spec_ref: "frontend_state_contract" +// @spec_ref frontend_state_contract export function $(selector: string, root: Document | HTMLElement = document): HTMLElement | null { return root.querySelector(selector) as HTMLElement | null; } +// @spec_ref frontend_state_contract export function on(el: HTMLElement | Document, type: K, handler: (ev: HTMLElementEventMap[K]) => void): void { el.addEventListener(type, handler as EventListener); } +// @spec_ref frontend_state_contract export function setText(el: HTMLElement | null, text: string): void { if (el) el.textContent = text; } +// @spec_ref frontend_state_contract export function setHTML(el: HTMLElement | null, html: string): void { if (el) el.innerHTML = html; } +// @spec_ref frontend_state_contract export function button(el: HTMLElement | null, disabled: boolean): void { if (el instanceof HTMLButtonElement) { el.disabled = disabled; diff --git a/utils/logger.js b/utils/logger.js index 3374dae..deaed7f 100644 --- a/utils/logger.js +++ b/utils/logger.js @@ -1,14 +1,19 @@ +// module_name: utils_logger_ts +// spec_ref: "frontend_state_contract" const ENABLE_LOG = true; +// @spec_ref frontend_state_contract export function log(scope, ...args) { if (!ENABLE_LOG) return; console.log(`[ChromeAI:${scope}]`, ...args); } +// @spec_ref frontend_state_contract export function warn(scope, ...args) { if (!ENABLE_LOG) return; console.warn(`[ChromeAI:${scope}]`, ...args); } +// @spec_ref frontend_state_contract export function error(scope, ...args) { console.error(`[ChromeAI:${scope}]`, ...args); } diff --git a/utils/logger.ts b/utils/logger.ts index 2f2514a..35cd1a2 100644 --- a/utils/logger.ts +++ b/utils/logger.ts @@ -1,15 +1,20 @@ +// module_name: utils_logger_ts +// spec_ref: "frontend_state_contract" const ENABLE_LOG = true; +// @spec_ref frontend_state_contract export function log(scope: string, ...args: unknown[]): void { if (!ENABLE_LOG) return; console.log(`[ChromeAI:${scope}]`, ...args); } +// @spec_ref frontend_state_contract export function warn(scope: string, ...args: unknown[]): void { if (!ENABLE_LOG) return; console.warn(`[ChromeAI:${scope}]`, ...args); } +// @spec_ref frontend_state_contract export function error(scope: string, ...args: unknown[]): void { console.error(`[ChromeAI:${scope}]`, ...args); } diff --git a/utils/storage.js b/utils/storage.js index 2284e5e..f42c5f6 100644 --- a/utils/storage.js +++ b/utils/storage.js @@ -1,3 +1,5 @@ +// module_name: utils_storage_ts +// spec_ref: "frontend_state_contract" import { error, log } from './logger.js'; const encoder = new TextEncoder(); const decoder = new TextDecoder(); diff --git a/utils/storage.ts b/utils/storage.ts index 4558746..8706359 100644 --- a/utils/storage.ts +++ b/utils/storage.ts @@ -1,3 +1,5 @@ +// module_name: utils_storage_ts +// spec_ref: "frontend_state_contract" import { error, log } from './logger.js'; const encoder = new TextEncoder(); diff --git a/utils/time.js b/utils/time.js index 4d8d8fc..3394af6 100644 --- a/utils/time.js +++ b/utils/time.js @@ -1,10 +1,15 @@ +// module_name: utils_time_ts +// spec_ref: "frontend_state_contract" +// @spec_ref frontend_state_contract export function nowISO() { return new Date().toISOString(); } +// @spec_ref frontend_state_contract export function daysFrom(timestamp) { const diff = Date.now() - timestamp; return Math.floor(diff / (1000 * 60 * 60 * 24)); } +// @spec_ref frontend_state_contract export function readableTime() { return new Date().toLocaleString(); } diff --git a/utils/time.ts b/utils/time.ts index 3d114cd..fc6e1da 100644 --- a/utils/time.ts +++ b/utils/time.ts @@ -1,12 +1,17 @@ +// module_name: utils_time_ts +// spec_ref: "frontend_state_contract" +// @spec_ref frontend_state_contract export function nowISO(): string { return new Date().toISOString(); } +// @spec_ref frontend_state_contract export function daysFrom(timestamp: number): number { const diff = Date.now() - timestamp; return Math.floor(diff / (1000 * 60 * 60 * 24)); } +// @spec_ref frontend_state_contract export function readableTime(): string { return new Date().toLocaleString(); } From 92e3b164cf65ccff9bef44fe4ebba7eed0b4e3cc Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 18 Aug 2026 21:26:31 +0200 Subject: [PATCH 14/66] docs(product): clarify tier and delivery boundaries Verified: pnpm lint && pnpm test --- README.md | 27 ++++++++++--------- ROADMAP.md | 77 ++++++++++++++++++++---------------------------------- 2 files changed, 42 insertions(+), 62 deletions(-) diff --git a/README.md b/README.md index 7c4391f..782288d 100644 --- a/README.md +++ b/README.md @@ -111,20 +111,21 @@ Your hardware and selected profile determine latency/quality. ## Tiers -### Core (Essential) -- Structured extraction -- Canonical metadata -- Manual export/copy +The repository configuration is authoritative for price and entitlement mapping: -### Connect (Plus) -- One-click connector exports -- Target-specific format adapters -- No persistent memory layer +| Tier | Price | Product boundary | +| --- | ---: | --- | +| Essential | $1.99 | Local structured extraction, canonical metadata and manual copy/export | +| Plus | $5.99 | Essential plus stateless summaries, batch clipping and connector-format exports | +| Pro | $14.99 | Plus plus multimodal processing and an explicit opt-in local knowledge layer | -### Knowledge (Deep) -- Explicit opt-in local memory layer -- Local embeddings/retrieval capabilities -- Inspect / export / delete retained knowledge +Essential and Plus do not retain a knowledge history. Pro stateful features must remain local, visible and user-controlled: retained data can be inspected, exported and deleted. A feature is available only when the signed entitlement and `pricing/tier-feature-map.json` both permit it; there is no silent downgrade or fallback. + +Prices and Paddle product identifiers live in `pricing/pricing-global.json` and `pricing/paddle-products.json`. Their presence is configuration evidence, not proof that production checkout or entitlement issuance is operational. + +### Team / self-hosted + +Team/self-hosted mode is planned, not shipped. Its admission contract requires an operator-owned Ollama endpoint, zero-access encrypted synchronization where synchronization is enabled, explicit tenant and retention controls, auditable entitlement administration, and the same no-cloud-fallback rule as the individual product. Until those paths pass real deployment, privacy, failure/recovery and multi-user isolation tests, SelectPilot makes no Team or self-hosted availability claim. --- @@ -157,7 +158,7 @@ Your hardware and selected profile determine latency/quality. Active development. Core local pipeline is functional and test-backed. -Current focus: stability, deterministic structure, and trust consistency. +Current focus: protected-main admission and physical Apple-hardware runtime verification. Team/self-hosted mode remains planned. --- diff --git a/ROADMAP.md b/ROADMAP.md index dfb7ecc..115b13c 100644 --- a/ROADMAP.md +++ b/ROADMAP.md @@ -1,6 +1,6 @@ # SelectPilot — Prioriterad åtgärdslista -> Harmoniserad med repo-läget per 2026-03-25 (granskning av `manifest.json`, `package.json`, `README.md`, teststruktur och `ZERO_LEAKAGE.md`). +> Harmoniserad med repo-, GitHub- och Linear-läget per 2026-08-18. > Prioritering: **P0** = blockerar trust/release · **P1** = produktkvalitet · **P2** = strategisk moat --- @@ -32,8 +32,8 @@ ### 4) Onboarding-friktion är för hög -- **Status:** 🟡 Delvis löst (bootstrap-script finns). -- **Åtgärd kvar:** ge ett verkligt one-command bootstrap-flöde som fungerar utan manuell felsökning. +- **Status:** ✅ Implementerat fail-closed; fysisk målmaskinsverifiering kvar. +- **Implementerat:** one-command bootstrap installerar/bygger, startar Ollama och bridge, väljer hårdvaruprofil, hämtar exakta modeller och avslutar först när generation- och embeddingkontrakten är friska. ### 5) Nginx/hosts-topologi bör vara osynlig @@ -75,8 +75,8 @@ ### 10) Generisk copilot-retorik underminerar edge -- **Status:** 🟡 Delvis adresserat. -- **Åtgärd kvar:** skärp positioning till structured extraction + local privacy i README/UI copy. +- **Status:** ✅ Löst i README. +- **Implementerat:** positionering och produktgräns utgår från deterministic structured extraction, canonical output och lokal privacy i stället för generisk copilot-retorik. ### 11) Presets är inte tillräckligt synliga eller utbyggbara @@ -85,14 +85,13 @@ ### 12) Ingen tydlig upgrade-path till team/self-hosted mode -- **Status:** 🟡 Planerad men ej konkretiserad. -- **Implementerat nu:** README förtydligar tier-guardrails där Essential/Plus hålls stateless i core-flödet och Pro definieras som explicit opt-in stateful lager. -- **Åtgärd kvar:** lägg kort sektion i README om Team/Self-hosted mode och avgränsa vad som är planerat. +- **Status:** ✅ Dokumentationskontrakt löst; funktionen är fortsatt uttryckligen planerad och inte skeppad. +- **Implementerat:** README avgränsar Team/Self-hosted mot operatorägd Ollama, zero-access-kryptering, tenant-/retentionkontroll, entitlementadministration och verifieringskrav före tillgänglighetsanspråk. ### 13) Tier-packaging och pricing-integritet behöver löpande styrning - **Status:** 🟡 Delvis adresserat. -- **Implementerat:** README innehåller nu explicit tier-prissättning, Paddle-mappning, positioneringsvarianter och realistisk ARR-prognos. +- **Implementerat:** README speglar repositoryts aktuella tier-priser, produktgränser och Paddle-mappning utan att beskriva konfiguration som fungerande produktionscheckout. Essential/Plus är stateless; Pro-state är explicit opt-in och användarstyrd. - **Åtgärd kvar:** - säkra att Pro kontinuerligt får tydlig premium-differentiering mot Plus - håll fast vid arkitekturkontraktet: stateless core i Essential/Plus, stateful funktioner endast via explicit opt-in i Pro @@ -104,22 +103,11 @@ ## Prioriterad genomförandeordning (från nu) -### Nästa sprint (måste först) - -1. **P1.7** Full extension E2E (markera text → sidepanel action → svar) -2. **P1.8** Strikt nätverksregressionstest för hela extension-flödet -3. **P1.6** Tydlig Quick Start (5 min) i README-toppen - -### Därefter (releasekvalitet) - -4. **P1.4** Minska onboarding-friktion ytterligare i bootstrap - -### Strategisk polish - -9. **P2.10** Positioning-copy: structured outputs + local-first -10. **P2.11** Utbyggbara presets (JSON/YAML + docs) -11. **P2.12** Team/Self-hosted mode i README -12. **P2.13** Tier-differentiering och pricing-integritet (särskilt Pro-value) +1. Behörig oberoende approval och skyddad merge av PR #22; auto-merge är aktivt och skydd får inte kringgås. +2. Verifiera exakt merge-SHA på `main` med full CI, E2E, privacy och säkerhetskontroller. +3. Kör bootstrap och verklig Ollama-inferens på M1 8GB och M4 16GB; bevara latency-, minnes-, lång-input-, failure- och recovery-evidens. +4. Verifiera eller håll M1 Max 32GB/Advanced-lanen explicit blockerad utan produktanspråk. +5. Slutbedöm hela systemet mot Linear-acceptans och Definition of Done; uppdatera issues/projekt först efter observerad evidens. --- @@ -130,18 +118,13 @@ ### 1) One-command installation (absolut först) - **Mål:** användaren ska klara onboarding med en enda kommando-rad. -- **Kvar att leverera:** - - gör `pnpm bootstrap:local` helt självbärande med robust felhantering - - automatisk kontroll av Ollama, modeller, launchd och `/health` - - tydlig slutrapport med exakt "klart/återstår" och nästa steg +- **Kodstatus:** implementerad med explicit felhantering och hälsokontroll för Ollama, exakta modeller, launchd och `/health`. +- **Kvar:** reproducerbar körning på avsedda Apple-målmaskiner. ### 2) "89-åring-läge" i README (ultrakort onboarding) - **Mål:** inga tekniska beslut i första flödet. -- **Kvar att leverera:** - - 3-stegs quick start högst upp i README (kopiera kommando → ladda extension → klart) - - separera nybörjarflöde från advanced/dev-sektioner - - lägg till "om något går fel" med copy/paste-kommandon +- **Status:** implementerad i README med tre steg och copy/paste-diagnostik. ### 3) UI/UX-polish till premium-kvalitet @@ -154,17 +137,13 @@ ### 4) E2E-bevis för verkligt användarflöde - **Mål:** kunna verifiera funktion och privacy med reproducerbara tester. -- **Kvar att leverera:** - - deterministisk start/stop av lokal server i testflödet - - full extension-path: markera text → sidepanel action → renderat svar - - strikt nätverksregression: ingen extern trafik i kärnflödet +- **Status:** automatiserat och grönt med deterministisk serverlivscykel, riktig unpacked-extension-path och strikt extern-nätverksblockering. +- **Kvar:** upprepa mot verklig Ollama på målmaskiner; testmocken är inte runtimebevis. ### 5) Strategisk finish (efter kvalitet + onboarding) -- **Kvar att leverera:** - - skärpt positioning-copy (structured extraction + local-first) - - presets som redigerbar JSON/YAML med dokumenterat schema - - tydlig Team/Self-hosted-sektion och migration path +- **Status:** positioning, redigerbara presets och Team/Self-hosted-gräns är dokumenterade. +- **Kvar:** inget Team/Self-hosted-tillgänglighetsanspråk får göras före en separat implementerad och verifierad vertikal. --- @@ -175,19 +154,19 @@ | 1 | Versionsdrift | P0 | Löst | Låg | | 2 | Dubbla lock-filer | P0 | Löst | Låg | | 3 | Ogiltiga manifest-fält | P0 | Löst | Låg | -| 4 | Onboarding-friktion | P1 | Delvis | Hög | +| 4 | Onboarding-friktion | P1 | Kod löst; fysisk verifiering kvar | Hög | | 5 | Nginx/hosts osynlighet | P1 | Löst* | Hög | -| 6 | README Quick Start | P1 | Delvis | Låg | -| 7 | E2E-tester extension | P1 | Delvis | Medium | -| 8 | Privacy regressionstester | P1 | Delvis | Medium | +| 6 | README Quick Start | P1 | Löst | Låg | +| 7 | E2E-tester extension | P1 | Löst | Medium | +| 8 | Privacy regressionstester | P1 | Löst | Medium | | 9 | Verifierbar privacy-yta | P1 | Löst | Medium | -| 10 | Omformulera positioning | P2 | Delvis | Låg | -| 11 | Utbyggbara presets | P2 | Delvis | Medium | -| 12 | Team/self-hosted-path | P2 | Delvis | Låg | +| 10 | Omformulera positioning | P2 | Löst | Låg | +| 11 | Utbyggbara presets | P2 | Löst | Medium | +| 12 | Team/self-hosted-kontrakt | P2 | Docs löst; produkt ej skeppad | Låg | | 13 | Tier/pricing-integritet | P2 | Delvis | Medium | \* Löst i huvudflödet, men behöver fortsatt skyddas från regression i docs/scripts. --- -*Senast uppdaterad: 2026-03-25* +*Senast uppdaterad: 2026-08-18* From 3f47e5ccd8bb16e2180178e621077d625e919bb2 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:26:15 +0200 Subject: [PATCH 15/66] ci: enforce exact-SHA admission evidence Expand required CI validation to cover privacy, unpacked-extension E2E, fail-closed policy compilation, and stress contracts. Emit an exact-SHA artifact that explicitly remains non-runtime and non-full-system evidence.\n\nVerified: pnpm lint && pnpm lint:manifest && pnpm typecheck && pnpm build && pnpm test && pnpm test:privacy && pnpm test:e2e && pnpm bench:frontier && pnpm runtime:compile-policy && pnpm runtime:validate-policy && node scripts/stress-runner.mjs && pnpm verify:admission --- .github/workflows/ci.yml | 37 ++++++++++ .gitignore | 1 + package.json | 1 + scripts/verify-code-admission.mjs | 113 ++++++++++++++++++++++++++++++ 4 files changed, 152 insertions(+) create mode 100644 scripts/verify-code-admission.mjs diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 9f0cd61..52cddc1 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -43,6 +43,9 @@ jobs: - name: Install dependencies run: pnpm install --frozen-lockfile + - name: Install Playwright Chromium + run: pnpm exec playwright install --with-deps chromium + - name: Lint run: pnpm lint @@ -63,3 +66,37 @@ jobs: - name: Test run: pnpm test + + - name: Verify privacy boundary + run: pnpm test:privacy + + - name: Verify unpacked extension E2E + run: pnpm test:e2e + + - name: Compile fail-closed benchmark evidence + run: pnpm bench:frontier + + - name: Compile runtime policy + run: pnpm runtime:compile-policy + + - name: Validate runtime policy + run: pnpm runtime:validate-policy + + - name: Verify deterministic stress contracts + run: node scripts/stress-runner.mjs + + - name: Verify exact-SHA code admission evidence + run: pnpm verify:admission + env: + SELECTPILOT_SOURCE_SHA: ${{ github.event.pull_request.head.sha || github.sha }} + SELECTPILOT_WORKFLOW_SHA: ${{ github.sha }} + SELECTPILOT_REF: ${{ github.ref }} + + - name: Upload exact-SHA code admission evidence + if: matrix.node-version == '22.x' + uses: actions/upload-artifact@v4 + with: + name: selectpilot-code-admission-${{ github.sha }} + path: reports/code_admission_evidence.json + if-no-files-found: error + retention-days: 30 diff --git a/.gitignore b/.gitignore index 5fdcbc0..91f43eb 100644 --- a/.gitignore +++ b/.gitignore @@ -5,5 +5,6 @@ package-lock.json npm-debug.log* test-results/ playwright-report/ +reports/code_admission_evidence.json /usr/local/var/log/chromeai/ /usr/local/var/run/chromeai/ diff --git a/package.json b/package.json index f402066..e0da878 100644 --- a/package.json +++ b/package.json @@ -23,6 +23,7 @@ "runtime:validate-policy": "node ./scripts/runtime-policy-validator.mjs", "runtime:rollback": "node ./scripts/runtime-rollback.mjs", "runtime:show-policy": "node ./scripts/runtime-show-policy.mjs", + "verify:admission": "node ./scripts/verify-code-admission.mjs", "sync:manifest-version": "node ./scripts/sync-manifest-version.mjs", "build": "pnpm sync:manifest-version && node ./scripts/generate-extraction-presets.mjs && tsc -p tsconfig.json && node ./scripts/build-extension.mjs", "lint": "node ./scripts/generate-extraction-presets.mjs --check && eslint --max-warnings=0 .", diff --git a/scripts/verify-code-admission.mjs b/scripts/verify-code-admission.mjs new file mode 100644 index 0000000..19c3cde --- /dev/null +++ b/scripts/verify-code-admission.mjs @@ -0,0 +1,113 @@ +#!/usr/bin/env node +// module_name: code_admission_verifier +// spec_ref: "verification" + +import fs from 'node:fs/promises'; +import path from 'node:path'; +import process from 'node:process'; +import { fileURLToPath } from 'node:url'; + +const repoRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); +const reportsDir = path.join(repoRoot, 'reports'); + +function fail(code, message, details = {}) { + process.stderr.write(`${JSON.stringify({ ok: false, code, message, details }, null, 2)}\n`); + process.exit(1); +} + +async function readJson(name) { + const filePath = path.join(reportsDir, name); + try { + return JSON.parse(await fs.readFile(filePath, 'utf8')); + } catch (error) { + fail('admission_evidence_unreadable', `Cannot read ${name}`, { error: String(error) }); + } +} + +function requireExactSha(value, field) { + if (!/^[0-9a-f]{40}$/i.test(value || '')) { + fail('admission_sha_invalid', `${field} must be an exact 40-character Git SHA`, { [field]: value || null }); + } + return value.toLowerCase(); +} + +const sourceSha = requireExactSha(process.env.SELECTPILOT_SOURCE_SHA, 'source_sha'); +const workflowSha = requireExactSha(process.env.SELECTPILOT_WORKFLOW_SHA, 'workflow_sha'); +const ref = process.env.SELECTPILOT_REF || ''; +if (!ref.startsWith('refs/')) fail('admission_ref_invalid', 'SELECTPILOT_REF must be an explicit Git ref', { ref }); + +const coverage = await readJson('spec_coverage_report.json'); +const verification = await readJson('verification_report.json'); +const policy = await readJson('../runtime/model_policy.json'); +const policyValidation = await readJson('runtime_policy_validation.json'); +const stress = await readJson('stress/master_summary.json'); + +const coverageSummary = { + implemented: coverage.implemented_spec_nodes?.length ?? 0, + missing: coverage.missing_spec_nodes?.length ?? -1, + unmapped_files: coverage.unmapped_files?.length ?? -1, + unmapped_functions: coverage.unmapped_functions?.length ?? -1, +}; +if ( + coverageSummary.implemented === 0 + || coverageSummary.missing !== 0 + || coverageSummary.unmapped_files !== 0 + || coverageSummary.unmapped_functions !== 0 +) { + fail('admission_traceability_incomplete', 'Specification traceability is incomplete', coverageSummary); +} + +if ( + verification.verification_scope !== 'static_traceability_only' + || verification.full_system_ok !== false + || verification.runtime_verified !== false +) { + fail('admission_scope_ambiguous', 'Verification evidence must remain explicitly scoped and fail closed', { + verification_scope: verification.verification_scope, + full_system_ok: verification.full_system_ok, + runtime_verified: verification.runtime_verified, + }); +} + +if ( + policy.promotion_evidence?.runtime_verified !== false + || policy.promotion_evidence?.status !== 'simulation_only_no_promotion' + || policy.promotion_history?.length !== 0 +) { + fail('admission_policy_promotes_unverified_evidence', 'Runtime policy accepted unverified promotion evidence', { + promotion_evidence: policy.promotion_evidence, + promotion_history_count: policy.promotion_history?.length ?? null, + }); +} + +if (policyValidation.ok !== true || stress.pass !== true) { + fail('admission_contract_validation_failed', 'Policy or stress validation is not green', { + policy_validation_ok: policyValidation.ok ?? null, + stress_pass: stress.pass ?? null, + }); +} + +const evidence = { + schema_version: 1, + evidence_class: 'exact_sha_code_admission', + source_sha: sourceSha, + workflow_sha: workflowSha, + ref, + generated_at: new Date().toISOString(), + checks: { + traceability: coverageSummary, + runtime_policy_validation: true, + deterministic_stress_contracts: true, + }, + runtime_verified: false, + physical_hardware_verified: false, + full_system_ok: false, + remaining_gates: [ + 'protected_main_merge', + 'physical_apple_hardware_runtime', + 'full_system_definition_of_done', + ], +}; + +await fs.writeFile(path.join(reportsDir, 'code_admission_evidence.json'), `${JSON.stringify(evidence, null, 2)}\n`); +process.stdout.write(`${JSON.stringify({ ok: true, output: 'reports/code_admission_evidence.json', ...evidence }, null, 2)}\n`); From 1662f161443541bf7df61d71b69fb2c2eb6ecca9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:28:29 +0200 Subject: [PATCH 16/66] ci: move actions to Node 24 runtimes Verified: workflow YAML parse and GitHub release metadata for action majors --- .github/workflows/ci.yml | 10 +++++----- .github/workflows/codeql.yml | 8 ++++---- 2 files changed, 9 insertions(+), 9 deletions(-) diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 52cddc1..55a7bb8 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -21,22 +21,22 @@ jobs: node-version: [20.x, 22.x] steps: - - uses: actions/checkout@v4 + - uses: actions/checkout@v7 - name: Use pnpm - uses: pnpm/action-setup@v4 + uses: pnpm/action-setup@v6 with: version: 10.33.0 - name: Use Node.js ${{ matrix.node-version }} - uses: actions/setup-node@v4 + uses: actions/setup-node@v7 with: node-version: ${{ matrix.node-version }} cache: pnpm cache-dependency-path: pnpm-lock.yaml - name: Use Python 3.12 - uses: actions/setup-python@v5 + uses: actions/setup-python@v7 with: python-version: "3.12" @@ -94,7 +94,7 @@ jobs: - name: Upload exact-SHA code admission evidence if: matrix.node-version == '22.x' - uses: actions/upload-artifact@v4 + uses: actions/upload-artifact@v7 with: name: selectpilot-code-admission-${{ github.sha }} path: reports/code_admission_evidence.json diff --git a/.github/workflows/codeql.yml b/.github/workflows/codeql.yml index fbf196c..d2d3785 100644 --- a/.github/workflows/codeql.yml +++ b/.github/workflows/codeql.yml @@ -27,7 +27,7 @@ jobs: language: ["javascript-typescript", "python"] steps: - - uses: actions/checkout@v4 + - uses: actions/checkout@v7 - name: Initialize CodeQL uses: github/codeql-action/init@v3 @@ -36,13 +36,13 @@ jobs: - name: Use pnpm if: matrix.language == 'javascript-typescript' - uses: pnpm/action-setup@v4 + uses: pnpm/action-setup@v6 with: version: 10.33.0 - name: Use Node.js 22 if: matrix.language == 'javascript-typescript' - uses: actions/setup-node@v4 + uses: actions/setup-node@v7 with: node-version: 22 cache: pnpm @@ -50,7 +50,7 @@ jobs: - name: Use Python 3.12 if: matrix.language == 'python' - uses: actions/setup-python@v5 + uses: actions/setup-python@v7 with: python-version: "3.12" From d7806fc0b42337b303f44b064a53f75718410eb2 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:31:49 +0200 Subject: [PATCH 17/66] fix(setup): repair model bootstrap contract Use the generated profile variable names consistently and add a side-effect-free plan mode with exact Fast and Balanced regression tests.\n\nVerified: bash -n scripts/bootstrap-macos-local.sh && pnpm lint && pnpm test && pnpm test:privacy && pnpm test:e2e --- scripts/bootstrap-macos-local.sh | 39 ++++++++++++++++--------- tests/panel/bootstrap-contract.test.mjs | 32 ++++++++++++++++++++ 2 files changed, 57 insertions(+), 14 deletions(-) create mode 100644 tests/panel/bootstrap-contract.test.mjs diff --git a/scripts/bootstrap-macos-local.sh b/scripts/bootstrap-macos-local.sh index 5c641f5..9969ae5 100755 --- a/scripts/bootstrap-macos-local.sh +++ b/scripts/bootstrap-macos-local.sh @@ -5,6 +5,7 @@ ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" PROFILE="auto" SKIP_OLLAMA_INSTALL="0" SKIP_MODEL_PULL="0" +PLAN_ONLY="0" BRIDGE_HEALTH_URL="http://127.0.0.1:8083/health" STATUS_OLLAMA_INSTALL="pending" @@ -27,6 +28,10 @@ while [[ $# -gt 0 ]]; do SKIP_MODEL_PULL="1" shift ;; + --plan) + PLAN_ONLY="1" + shift + ;; *) echo "Unknown argument: $1" >&2 exit 1 @@ -34,11 +39,6 @@ while [[ $# -gt 0 ]]; do esac done -if [[ "$(uname -s)" != "Darwin" ]]; then - echo "This bootstrapper currently supports macOS only." >&2 - exit 1 -fi - read_profile_json() { python3 - "$PROFILE" "$ROOT" <<'PY' import json @@ -55,11 +55,12 @@ recommendation = recommend_runtime_profile() selected = recommendation["recommended_profile"] if profile == "auto" else profile runtime_profile = get_runtime_profile(selected) commands = build_bootstrap_commands(runtime_profile.key, root) +reason = recommendation["reason"] if profile == "auto" else "Explicit profile selected by operator." print(json.dumps({ "selected_profile": runtime_profile.key, "label": runtime_profile.label, - "reason": recommendation["reason"], + "reason": reason, "generation_model": runtime_profile.generation_model, "embedding_model": runtime_profile.embedding_model, "num_ctx": runtime_profile.num_ctx, @@ -81,6 +82,16 @@ PY )" eval "$PROFILE_VARS" +if [[ "$PLAN_ONLY" == "1" ]]; then + printf '%s\n' "$PROFILE_JSON" + exit 0 +fi + +if [[ "$(uname -s)" != "Darwin" ]]; then + echo "This bootstrapper currently supports macOS only." >&2 + exit 1 +fi + if [[ "$SKIP_OLLAMA_INSTALL" != "1" ]] && ! command -v ollama >/dev/null 2>&1; then if command -v brew >/dev/null 2>&1; then echo "Installing Ollama with Homebrew..." @@ -118,17 +129,17 @@ if [[ "$STATUS_OLLAMA_RUNNING" != "ok" ]]; then fi if [[ "$SKIP_MODEL_PULL" != "1" ]]; then - echo "Pulling generation model: $GEN_MODEL" - ollama pull "$GEN_MODEL" - echo "Pulling embedding model: $EMBED_MODEL" - ollama pull "$EMBED_MODEL" + echo "Pulling generation model: $GENERATION_MODEL" + ollama pull "$GENERATION_MODEL" + echo "Pulling embedding model: $EMBEDDING_MODEL" + ollama pull "$EMBEDDING_MODEL" STATUS_MODEL_PULL="ok" else STATUS_MODEL_PULL="skipped" fi -CHROMEAI_OLLAMA_MODEL="$GEN_MODEL" \ -CHROMEAI_OLLAMA_EMBED_MODEL="$EMBED_MODEL" \ +CHROMEAI_OLLAMA_MODEL="$GENERATION_MODEL" \ +CHROMEAI_OLLAMA_EMBED_MODEL="$EMBEDDING_MODEL" \ CHROMEAI_OLLAMA_NUM_CTX="$NUM_CTX" \ "$ROOT/scripts/install-macos-local.sh" STATUS_LAUNCHAGENT="ok" @@ -166,8 +177,8 @@ SelectPilot bootstrap complete. Profile: $SELECTED_PROFILE Reason: $REASON -Generation model: $GEN_MODEL -Embedding model: $EMBED_MODEL +Generation model: $GENERATION_MODEL +Embedding model: $EMBEDDING_MODEL Context window: $NUM_CTX Next recommended command: diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs new file mode 100644 index 0000000..cc26538 --- /dev/null +++ b/tests/panel/bootstrap-contract.test.mjs @@ -0,0 +1,32 @@ +// module_name: bootstrap_contract_test +// spec_ref: "runtime_profiles" + +import assert from 'node:assert/strict'; +import { execFileSync } from 'node:child_process'; +import test from 'node:test'; + +function bootstrapPlan(profile) { + const stdout = execFileSync('bash', ['./scripts/bootstrap-macos-local.sh', '--profile', profile, '--plan'], { + cwd: process.cwd(), + encoding: 'utf8', + }); + return JSON.parse(stdout); +} + +test('bootstrap plan exposes exact fail-closed Fast runtime contract without side effects', () => { + const plan = bootstrapPlan('fast'); + assert.equal(plan.selected_profile, 'fast'); + assert.equal(plan.generation_model, 'gemma4:e2b-it-qat'); + assert.equal(plan.embedding_model, 'nomic-embed-text-v2-moe:latest'); + assert.equal(plan.num_ctx, 16_384); + assert.equal(plan.reason, 'Explicit profile selected by operator.'); +}); + +test('bootstrap plan exposes exact fail-closed Balanced runtime contract without side effects', () => { + const plan = bootstrapPlan('balanced'); + assert.equal(plan.selected_profile, 'balanced'); + assert.equal(plan.generation_model, 'gemma4:e4b-it-qat'); + assert.equal(plan.embedding_model, 'nomic-embed-text-v2-moe:latest'); + assert.equal(plan.num_ctx, 32_768); + assert.equal(plan.reason, 'Explicit profile selected by operator.'); +}); From e08157f7b1bb31e4869dfe700a5f056194ce6404 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:33:53 +0200 Subject: [PATCH 18/66] fix(setup): honor automatic hardware selection Verified: node --test tests/panel/bootstrap-contract.test.mjs --- package.json | 2 +- tests/panel/bootstrap-contract.test.mjs | 6 ++++++ 2 files changed, 7 insertions(+), 1 deletion(-) diff --git a/package.json b/package.json index e0da878..7a1853d 100644 --- a/package.json +++ b/package.json @@ -10,7 +10,7 @@ }, "description": "Privacy-first local execution layer for selected text, powered by Ollama.", "scripts": { - "setup:local": "pnpm install && pnpm build && bash ./scripts/bootstrap-macos-local.sh --profile fast", + "setup:local": "pnpm install && pnpm build && bash ./scripts/bootstrap-macos-local.sh --profile auto", "bootstrap:local": "bash ./scripts/bootstrap-macos-local.sh", "benchmark:local": "bash ./scripts/benchmark-local-runtime.sh", "spec:compile": "node ./scripts/spec-compiler.mjs", diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index cc26538..fb31f72 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -3,6 +3,7 @@ import assert from 'node:assert/strict'; import { execFileSync } from 'node:child_process'; +import { readFileSync } from 'node:fs'; import test from 'node:test'; function bootstrapPlan(profile) { @@ -30,3 +31,8 @@ test('bootstrap plan exposes exact fail-closed Balanced runtime contract without assert.equal(plan.num_ctx, 32_768); assert.equal(plan.reason, 'Explicit profile selected by operator.'); }); + +test('one-command setup preserves hardware-aware automatic profile selection', () => { + const packageJson = JSON.parse(readFileSync('./package.json', 'utf8')); + assert.match(packageJson.scripts['setup:local'], /bootstrap-macos-local\.sh --profile auto$/); +}); From acd636721b467b9f14a8eab9d17e752a4c7f314b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:34:20 +0200 Subject: [PATCH 19/66] ci: lint shell execution contracts Verified: shellcheck scripts/*.sh on Hetzner --- .github/workflows/ci.yml | 3 +++ 1 file changed, 3 insertions(+) diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 55a7bb8..c4226ab 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -52,6 +52,9 @@ jobs: - name: Lint manifest run: pnpm lint:manifest + - name: Lint shell contracts + run: shellcheck scripts/*.sh + - name: Sync manifest version run: pnpm sync:manifest-version From e4b75cd26ea1d5ff6b65deaa921af9fb5a8e4327 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:38:38 +0200 Subject: [PATCH 20/66] fix(runtime): fail closed on unavailable models Remove synthetic selection when no declared model is installed, keep declared fallback behavior observable, and report promotion only for runtime-verified matching history.\n\nVerified: shellcheck scripts/*.sh && pnpm lint && pnpm typecheck && pnpm build && pnpm test && pnpm test:privacy && pnpm test:e2e && pnpm bench:frontier && pnpm runtime:compile-policy && pnpm runtime:validate-policy && node scripts/stress-runner.mjs && pnpm verify:admission --- scripts/stress-runner.mjs | 8 ++- server/model/runtimeSelectorAdapter.js | 20 ++---- server/model/runtimeSelectorAdapter.ts | 24 +++---- server/nano_server.py | 13 +++- shared/types/runtimePolicy.ts | 4 ++ tests/panel/runtime-selector.test.mjs | 84 ++++++++++++++++++++++++ tests/server/test_validation_pipeline.py | 39 +++++++++++ 7 files changed, 162 insertions(+), 30 deletions(-) create mode 100644 tests/panel/runtime-selector.test.mjs diff --git a/scripts/stress-runner.mjs b/scripts/stress-runner.mjs index c80ec1d..d31b6c6 100644 --- a/scripts/stress-runner.mjs +++ b/scripts/stress-runner.mjs @@ -428,7 +428,13 @@ async function runSingleTest(test, monolith, bindings, runtimePolicy, runtimeReg const installedModelIds = (runtimeRegistry.models || []) .filter((model) => model.installation_state === 'installed') .map((model) => model.model_id); - const availableModelIds = expectedRuntimeMismatchAllowed ? [] : installedModelIds; + const matchingPolicyTuple = (runtimePolicy.defaults || []).find((entry) => + entry.task_family === taskAnalysis.task_family + && entry.output_mode === taskAnalysis.output_mode + && entry.hardware_profile === taskAnalysis.hardware_profile); + const availableModelIds = expectedRuntimeMismatchAllowed + ? installedModelIds.filter((modelId) => modelId !== matchingPolicyTuple?.preferred_model_id) + : installedModelIds; const selected = bindings.selectRuntimeModel( { diff --git a/server/model/runtimeSelectorAdapter.js b/server/model/runtimeSelectorAdapter.js index 8c42272..7e5336d 100644 --- a/server/model/runtimeSelectorAdapter.js +++ b/server/model/runtimeSelectorAdapter.js @@ -3,6 +3,10 @@ export function selectRuntimeModel(input, policy, registry) { const available = new Set(input.availableModelIds); const registryById = new Map(registry.models.map((m) => [m.model_id, m])); const quarantined = new Set((policy.quarantined_models || []).map((q) => q.model_id)); + const isVerifiedPromotion = (modelId) => policy.promotion_evidence?.runtime_verified === true + && (policy.promotion_history || []).some((entry) => entry.task_family === input.taskFamily + && entry.hardware_profile === input.hardwareProfile + && entry.new_model_id === modelId); const isSelectable = (modelId, allowQuarantined = false) => { if (!available.has(modelId)) return false; @@ -33,28 +37,18 @@ export function selectRuntimeModel(input, policy, registry) { selection_path: 'runtime_policy_preferred', selection_reason: tuple.selection_reason, policy_version: policy.policy_version, - promotion_applied: true, + promotion_applied: isVerifiedPromotion(tuple.preferred_model_id), }; } const fallback = (tuple.fallback_model_ids || []).find((id) => isSelectable(id)); - if (!fallback) { - if (input.availableModelIds.length === 0) { - return { - selected_model_id: tuple.preferred_model_id, - selection_path: 'runtime_policy_fallback', - selection_reason: 'runtime_policy_fallback_no_models_available_simulated_degraded_mode', - policy_version: policy.policy_version, - promotion_applied: false, - }; - } + if (!fallback) return null; - } return { selected_model_id: fallback, selection_path: 'runtime_policy_fallback', selection_reason: 'runtime_policy_fallback_models_in_order', policy_version: policy.policy_version, - promotion_applied: true, + promotion_applied: false, }; } // module_name: runtime_model_selector diff --git a/server/model/runtimeSelectorAdapter.ts b/server/model/runtimeSelectorAdapter.ts index 19be589..1e20dc6 100644 --- a/server/model/runtimeSelectorAdapter.ts +++ b/server/model/runtimeSelectorAdapter.ts @@ -21,6 +21,13 @@ export function selectRuntimeModel( const registryById = new Map(registry.models.map((m) => [m.model_id, m])); const quarantined = new Set((policy.quarantined_models || []).map((q) => q.model_id)); + const isVerifiedPromotion = (modelId: string) => + policy.promotion_evidence?.runtime_verified === true + && (policy.promotion_history || []).some((entry) => + entry.task_family === input.taskFamily + && entry.hardware_profile === input.hardwareProfile + && entry.new_model_id === modelId); + const isSelectable = (modelId: string, allowQuarantined = false) => { if (!available.has(modelId)) return false; if (!registryById.has(modelId)) return false; @@ -52,30 +59,19 @@ export function selectRuntimeModel( selection_path: 'runtime_policy_preferred', selection_reason: tuple.selection_reason, policy_version: policy.policy_version, - promotion_applied: true, + promotion_applied: isVerifiedPromotion(tuple.preferred_model_id), }; } const fallback = (tuple.fallback_model_ids || []).find((id) => isSelectable(id)); - if (!fallback) { - if (input.availableModelIds.length === 0) { - return { - selected_model_id: tuple.preferred_model_id, - selection_path: 'runtime_policy_fallback', - selection_reason: 'runtime_policy_fallback_no_models_available_simulated_degraded_mode', - policy_version: policy.policy_version, - promotion_applied: false, - }; - } - return null; - } + if (!fallback) return null; return { selected_model_id: fallback, selection_path: 'runtime_policy_fallback', selection_reason: 'runtime_policy_fallback_models_in_order', policy_version: policy.policy_version, - promotion_applied: true, + promotion_applied: false, }; } // module_name: runtime_model_selector diff --git a/server/nano_server.py b/server/nano_server.py index 5ecc724..e45a2cd 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -460,6 +460,8 @@ def _runtime_policy_select( return None policy_version = str(policy.get("policy_version") or "") or None + promotion_evidence = policy.get("promotion_evidence") if isinstance(policy.get("promotion_evidence"), dict) else {} + promotion_history = policy.get("promotion_history") if isinstance(policy.get("promotion_history"), list) else [] defaults = policy.get("defaults") if isinstance(policy.get("defaults"), list) else [] quarantined = set() for item in policy.get("quarantined_models") or []: @@ -519,13 +521,20 @@ def _model_available(model_id: str) -> bool: chosen = matches[0] preferred = str(chosen.get("preferred_model_id") or "").strip() + verified_preferred_promotion = bool(promotion_evidence.get("runtime_verified")) and any( + isinstance(item, dict) + and str(item.get("task_family") or "") == task_family + and str(item.get("hardware_profile") or "") == hardware_profile + and str(item.get("new_model_id") or "") == preferred + for item in promotion_history + ) if preferred and _model_available(preferred) and _hardware_allows(preferred) and preferred not in quarantined and not _is_quarantined(preferred): return { "model_id": preferred, "selection_path": "runtime_policy_preferred", "selection_reason": str(chosen.get("selection_reason") or "runtime_policy_preferred_model_if_available_and_not_quarantined"), "policy_version": policy_version, - "promotion_applied": True, + "promotion_applied": verified_preferred_promotion, } for fallback in chosen.get("fallback_model_ids") or []: @@ -542,7 +551,7 @@ def _model_available(model_id: str) -> bool: "selection_path": "runtime_policy_fallback", "selection_reason": "runtime_policy_fallback_models_in_order", "policy_version": policy_version, - "promotion_applied": True, + "promotion_applied": False, } return None diff --git a/shared/types/runtimePolicy.ts b/shared/types/runtimePolicy.ts index 45ff45e..23d0fdc 100644 --- a/shared/types/runtimePolicy.ts +++ b/shared/types/runtimePolicy.ts @@ -32,6 +32,10 @@ export type RuntimePolicyHistory = { export type RuntimeModelPolicy = { policy_version: string; generated_at_unix_ms: number; + promotion_evidence?: { + runtime_verified: boolean; + status: string; + }; source_reports: string[]; global_guards: { determinism_min: number; diff --git a/tests/panel/runtime-selector.test.mjs b/tests/panel/runtime-selector.test.mjs new file mode 100644 index 0000000..1e09354 --- /dev/null +++ b/tests/panel/runtime-selector.test.mjs @@ -0,0 +1,84 @@ +// module_name: runtime_selector_test +// spec_ref: "model_selection_layer" + +import assert from 'node:assert/strict'; +import test from 'node:test'; + +import { selectRuntimeModel } from '../../server/model/runtimeSelectorAdapter.js'; + +const policy = { + policy_version: 'test-policy', + promotion_evidence: { runtime_verified: false, status: 'simulation_only_no_promotion' }, + promotion_history: [], + quarantined_models: [], + defaults: [{ + task_family: 'extract', + output_mode: 'strict_json', + hardware_profile: 'medium', + preferred_model_id: 'model:preferred', + fallback_model_ids: ['model:fallback'], + selection_reason: 'test_preferred', + }], +}; + +const registry = { + models: [ + { model_id: 'model:preferred' }, + { model_id: 'model:fallback' }, + ], +}; + +function select(availableModelIds) { + return selectRuntimeModel({ + taskFamily: 'extract', + outputMode: 'strict_json', + hardwareProfile: 'medium', + availableModelIds, + }, policy, registry); +} + +test('runtime selection fails closed when no declared model is installed', () => { + assert.equal(select([]), null); +}); + +test('runtime selection uses the exact preferred model when installed', () => { + assert.deepEqual(select(['model:preferred']), { + selected_model_id: 'model:preferred', + selection_path: 'runtime_policy_preferred', + selection_reason: 'test_preferred', + policy_version: 'test-policy', + promotion_applied: false, + }); +}); + +test('runtime selection exposes an explicitly declared fallback path', () => { + assert.deepEqual(select(['model:fallback']), { + selected_model_id: 'model:fallback', + selection_path: 'runtime_policy_fallback', + selection_reason: 'runtime_policy_fallback_models_in_order', + policy_version: 'test-policy', + promotion_applied: false, + }); +}); + +test('runtime selection reports promotion only for verified matching history', () => { + const promotedPolicy = { + ...policy, + promotion_evidence: { runtime_verified: true, status: 'runtime_verified' }, + promotion_history: [{ + task_family: 'extract', + hardware_profile: 'medium', + previous_model_id: 'model:old', + new_model_id: 'model:preferred', + decision_reason: 'runtime_evidence', + effective_from_unix_ms: 1, + }], + }; + const selected = selectRuntimeModel({ + taskFamily: 'extract', + outputMode: 'strict_json', + hardwareProfile: 'medium', + availableModelIds: ['model:preferred'], + }, promotedPolicy, registry); + assert.equal(selected?.promotion_applied, true); +}); diff --git a/tests/server/test_validation_pipeline.py b/tests/server/test_validation_pipeline.py index 4b7af36..5bfd668 100644 --- a/tests/server/test_validation_pipeline.py +++ b/tests/server/test_validation_pipeline.py @@ -17,6 +17,7 @@ build_runtime_meta_event, compile_intent_to_ir, ValidationError, + _runtime_policy_select, _resolve_trace_id, enforce_contract_fields, get_operation_contract, @@ -34,6 +35,44 @@ class ValidationPipelineTests(unittest.TestCase): + def test_runtime_policy_does_not_claim_unverified_promotion(self) -> None: + policy = { + "policy_version": "test-policy", + "promotion_evidence": {"runtime_verified": False, "status": "simulation_only_no_promotion"}, + "promotion_history": [], + "quarantined_models": [], + "defaults": [{ + "task_family": "extract", + "output_mode": "strict_json", + "hardware_profile": "medium", + "preferred_model_id": "model:preferred", + "fallback_model_ids": ["model:fallback"], + "selection_reason": "baseline_registry_no_runtime_promotion", + }], + } + registry = { + "models": [ + {"model_id": "model:preferred", "min_hardware_profile": "low"}, + {"model_id": "model:fallback", "min_hardware_profile": "low"}, + ], + } + task = {"task_type": "extract", "output_structure": "strict_json", "hardware_profile": "medium"} + + with patch("nano_server._load_json_file", side_effect=[policy, registry]), patch( + "nano_server._is_quarantined", return_value=False + ): + preferred = _runtime_policy_select(task_analysis=task, available_model_ids=["model:preferred"]) + self.assertIsNotNone(preferred) + self.assertFalse(preferred["promotion_applied"]) + + with patch("nano_server._load_json_file", side_effect=[policy, registry]), patch( + "nano_server._is_quarantined", return_value=False + ): + fallback = _runtime_policy_select(task_analysis=task, available_model_ids=["model:fallback"]) + self.assertIsNotNone(fallback) + self.assertEqual(fallback["selection_path"], "runtime_policy_fallback") + self.assertFalse(fallback["promotion_applied"]) + def test_summarize_payload_requires_non_empty_text(self) -> None: with self.assertRaises(ValidationError) as ctx: validate_summarize_payload({"text": " "}) From ec809fc0a1502a0740204994b5b21244c4436d76 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:42:13 +0200 Subject: [PATCH 21/66] docs(runtime): restore canonical lane contract Verified: repository link and Linear authority reconciliation --- README.md | 2 + docs/AI_RUNTIME_LANE_MATRIX.md | 76 ++++++++++++++++++++++++++++++++++ 2 files changed, 78 insertions(+) create mode 100644 docs/AI_RUNTIME_LANE_MATRIX.md diff --git a/README.md b/README.md index 782288d..4e2e8de 100644 --- a/README.md +++ b/README.md @@ -83,6 +83,8 @@ Ollama (local models) - No external inference endpoints on core path - Privacy boundary is observable and testable +Runtime authority and candidate-lane promotion rules are defined in [`docs/AI_RUNTIME_LANE_MATRIX.md`](docs/AI_RUNTIME_LANE_MATRIX.md). + --- ## Privacy Model diff --git a/docs/AI_RUNTIME_LANE_MATRIX.md b/docs/AI_RUNTIME_LANE_MATRIX.md new file mode 100644 index 0000000..25a66e6 --- /dev/null +++ b/docs/AI_RUNTIME_LANE_MATRIX.md @@ -0,0 +1,76 @@ +# SelectPilot AI Runtime Lane Matrix + +This contract defines which local runtime may perform each class of work without changing output authority, privacy, or user control. + +## Current truth + +- Ollama is the only implemented and authoritative inference lane. +- Gemini Nano and WebGPU are candidates, not current product capabilities. +- Candidate lanes remain inactive until capability, privacy, benchmark, failure/recovery, and contract evidence passes on the real unpacked-extension path. +- An unavailable or unverified lane fails closed. There is no silent failover to another provider. + +## Decision rule + +1. Output entering Golden JSON uses Ollama. +2. Optional, replaceable, non-authoritative assistance may become eligible for Gemini Nano after its promotion gate passes. +3. Semantics-preserving acceleration may become eligible for WebGPU after its promotion gate passes. +4. Unproven availability means inactive. + +## Lane matrix + +| Operation | Ollama | Gemini Nano | WebGPU | +| --- | --- | --- | --- | +| Structured extraction | Required | Forbidden | Support only | +| Canonical normalization | Required | Forbidden | Support only | +| Preset execution | Required | Forbidden | Support only | +| Golden JSON summary | Required | Forbidden | Support only | +| Disposable preview | Available | Eligible after promotion | Support only | +| Rewrite suggestion | Available | Eligible after promotion | Support only | +| Intent or preset hint | Available | Eligible after promotion | Support only | +| Deterministic pre/post-processing | Available | Forbidden when canonical | Eligible after promotion | +| Runtime benchmarking | Measured on physical target | Measured only when enabled | Measured only when enabled | + +“Support only” means acceleration without semantic authority. It cannot create, repair, replace, or silently alter canonical output. + +## Hard boundaries + +- No silent provider or model failover. +- No remote inference fallback. +- Gemini Nano cannot create, repair, or persist Golden JSON. +- WebGPU is acceleration, never semantic authority. +- Capability detection sends no selected text and creates no model session. +- Runtime identity, locality, selected profile, context window, failure, and fallback path remain observable. +- A declared model fallback must be installed, policy-allowed, non-quarantined, and surfaced as a fallback. +- Simulation evidence cannot promote a lane or model. + +## Promotion gate + +A candidate lane may ship only when all of the following are true: + +- implemented end to end for its allowed operation set; +- privacy and strict no-external-network tests pass; +- success, failure, and recovery paths pass; +- repeatable benefit is measured on every supported hardware class; +- output equivalence is proven, or the output is structurally non-authoritative; +- UI truth identifies runtime, locality, and execution state without ambiguity; +- evidence is bound to an exact protected-main SHA; +- Linear acceptance and dependencies are reconciled. + +Until every gate passes, the lane remains inactive and must not appear as an available product capability. + +## Evidence classes + +| Evidence | Proves | Does not prove | +| --- | --- | --- | +| Static traceability | Source-to-contract mapping | Runtime behavior | +| Deterministic simulation | Repeatable policy/scenario logic | Hardware performance or promotion eligibility | +| CI/E2E code admission | Exact code path passes automated checks | Physical local-model execution | +| Physical runtime evidence | Observed behavior on a named target and exact SHA | Other hardware classes | +| Full-system DoD | All protected delivery and runtime gates pass | Nothing beyond the declared release scope | + +## Source authority + +- This file owns the versioned runtime-lane contract. +- Linear owns planning, status, dependencies, acceptance, and evidence. +- The Linear architecture resource is [SelectPilot AI Runtime Lane Matrix](https://linear.app/sodersten-space/document/selectpilot-ai-runtime-lane-matrix-aecba24119c7). +- Conflicts fail closed until repository and Linear sources are reconciled. From 0259a2fa1d3d23e06cc4ea8413453f65d548b579 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 00:43:43 +0200 Subject: [PATCH 22/66] fix(entitlements): fail closed on token verification Route entitlement checks to the explicit local billing authority and reset cached privileges whenever a token changes. Verified: pnpm lint, pnpm typecheck, pnpm test, pnpm build, pnpm test:privacy, pnpm test:e2e, pnpm bench:frontier, pnpm runtime:compile-policy, pnpm runtime:validate-policy, node scripts/stress-runner.mjs --- background/entitlement-service.ts | 20 +-- server/nano_server.py | 133 ++++++------------ tests/server/test_entitlement_verification.py | 70 +++++++++ 3 files changed, 124 insertions(+), 99 deletions(-) create mode 100644 tests/server/test_entitlement_verification.py diff --git a/background/entitlement-service.ts b/background/entitlement-service.ts index da2b303..1a3de4a 100644 --- a/background/entitlement-service.ts +++ b/background/entitlement-service.ts @@ -223,17 +223,17 @@ async function remoteVerify(token: string): Promise { } export async function setEntitlementToken(token: string): Promise { - const existing = await loadLicense(); + const normalizedToken = token.trim(); + if (!normalizedToken) throw new Error('entitlement_token_required'); + + const resetAt = nowMs(); await saveLicense({ - token, - tier: existing?.tier || 'essential', - issuedAt: existing?.issuedAt || nowMs(), - expiresAt: existing?.expiresAt, - features: existing?.features, - cachedAt: existing?.cachedAt, - signature: existing?.signature, - alg: existing?.alg, - kid: existing?.kid, + token: normalizedToken, + tier: 'essential', + issuedAt: resetAt, + expiresAt: resetAt, + features: [], + cachedAt: 0, }); } diff --git a/server/nano_server.py b/server/nano_server.py index e45a2cd..b491679 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -6,8 +6,6 @@ # @spec_ref failure_isolation # @spec_ref latency_budget import argparse -import base64 -import hmac import hashlib import json import os @@ -20,6 +18,8 @@ from pathlib import Path from datetime import datetime, timezone from urllib.parse import urlparse, parse_qs +from urllib.error import HTTPError, URLError +from urllib.request import Request, urlopen from ollama_client import OllamaClient, OllamaError from extraction_presets import get_extraction_preset @@ -1276,94 +1276,49 @@ def benchmark_runtime() -> dict: def license_verify(payload: dict) -> dict: - token = payload.get("token", "") - tier = "pro" if "pro" in token else "plus" if "plus" in token else "essential" - token_basis = str(token or "") - token_hash = hashlib.sha256(token_basis.encode("utf-8")).hexdigest() - now = 1_700_000_000_000 + int(token_hash[:8], 16) - features_by_tier = { - "essential": [ - "selection_clipping", - "markdown_export", - "clipboard_export", - "side_panel_ui", - "structured_extraction", - "canonical_metadata", - "local_processing", - ], - "plus": [ - "selection_clipping", - "markdown_export", - "clipboard_export", - "side_panel_ui", - "structured_extraction", - "canonical_metadata", - "local_processing", - "text_summarization", - "basic_local_agent", - "export_obsidian", - "export_notion", - "export_mem_ai", - "export_apple_notes", - "format_adapters", - "one_click_export", - "batch_clipping", - "structured_summaries", - ], - "pro": [ - "selection_clipping", - "markdown_export", - "clipboard_export", - "side_panel_ui", - "structured_extraction", - "canonical_metadata", - "local_processing", - "text_summarization", - "basic_local_agent", - "export_obsidian", - "export_notion", - "export_mem_ai", - "export_apple_notes", - "format_adapters", - "one_click_export", - "batch_clipping", - "structured_summaries", - "audio_transcription", - "video_frame_ocr", - "image_ocr", - "multimodal_clipper", - "local_embeddings", - "advanced_agent_reasoning", - "project_memory", - "knowledge_graph", - "offline_search", - "auto_history_indexing", - ], - } - entitlement = { - "token": token, - "tier": tier, - "features": features_by_tier.get(tier, []), - "issuedAt": now, - "expiresAt": now + 30 * 24 * 60 * 60 * 1000, - } + token = str(payload.get("token") or "").strip() + if not token: + raise ValidationError("missing_token", "Entitlement token is required") - # Minimal signed payload support for MVP environments. - # In production, replace with Ed25519 signing and public-key verification in client. - signing_secret = os.environ.get("CHROMEAI_ENTITLEMENT_SIGNING_SECRET", "") - canonical = json.dumps(entitlement, separators=(",", ":"), ensure_ascii=False) - signature = "" - if signing_secret: - digest = hmac.new(signing_secret.encode("utf-8"), canonical.encode("utf-8"), hashlib.sha256).digest() - signature = base64.b64encode(digest).decode("ascii") - - response = { - "entitlement": entitlement, - "signature": signature, - "alg": "HMAC-SHA256" if signing_secret else "none", - "kid": "local-dev", - } - return response + verifier_url = os.environ.get( + "SELECTPILOT_BILLING_VERIFY_URL", + "http://127.0.0.1:8090/license/verify", + ) + parsed = urlparse(verifier_url) + if parsed.scheme != "http" or parsed.hostname not in LOCAL_HOSTS or parsed.path != "/license/verify": + raise ValidationError( + "invalid_entitlement_verifier", + "Entitlement verifier must be the explicit local HTTP license endpoint", + status=503, + ) + + request = Request( + verifier_url, + data=json.dumps({"token": token}, separators=(",", ":")).encode("utf-8"), + headers={"Content-Type": "application/json"}, + method="POST", + ) + try: + with urlopen(request, timeout=2.0) as response: + result = json.loads(response.read().decode("utf-8")) + except HTTPError as exc: + if exc.code == 401: + raise ValidationError("invalid_entitlement", "Entitlement token is invalid", status=401) from exc + raise ValidationError("entitlement_verifier_error", "Local entitlement verifier rejected the request", status=503) from exc + except (URLError, TimeoutError, json.JSONDecodeError) as exc: + raise ValidationError("entitlement_verifier_unavailable", "Local entitlement verifier is unavailable", status=503) from exc + + issued_at = result.get("issuedAt") if isinstance(result, dict) else None + expires_at = result.get("expiresAt") if isinstance(result, dict) else None + if ( + not isinstance(result, dict) + or result.get("tier") not in {"essential", "plus", "pro"} + or not isinstance(issued_at, int) + or isinstance(issued_at, bool) + or (expires_at is not None and (not isinstance(expires_at, int) or isinstance(expires_at, bool))) + ): + raise ValidationError("invalid_entitlement_response", "Local entitlement verifier returned an invalid contract", status=503) + return result class Handler(BaseHTTPRequestHandler): diff --git a/tests/server/test_entitlement_verification.py b/tests/server/test_entitlement_verification.py new file mode 100644 index 0000000..21bd2fa --- /dev/null +++ b/tests/server/test_entitlement_verification.py @@ -0,0 +1,70 @@ +"""module_name: entitlement_verification_tests; spec_ref: "testing_strategy.integration_tests".""" + +from __future__ import annotations + +import io +import os +import sys +import unittest +from pathlib import Path +from unittest.mock import patch +from urllib.error import HTTPError, URLError + +ROOT = Path(__file__).resolve().parents[2] +SERVER_DIR = ROOT / "server" +if str(SERVER_DIR) not in sys.path: + sys.path.insert(0, str(SERVER_DIR)) + +from nano_server import ValidationError, license_verify # noqa: E402 + + +class _Response: + def __init__(self, body: bytes): + self.body = body + + def __enter__(self): + return self + + def __exit__(self, *_args): + return False + + def read(self) -> bytes: + return self.body + + +class EntitlementVerificationTests(unittest.TestCase): + def test_pro_substring_cannot_grant_entitlement_without_authority(self) -> None: + with patch("nano_server.urlopen", side_effect=URLError("offline")): + with self.assertRaises(ValidationError) as ctx: + license_verify({"token": "anything-pro-anything"}) + self.assertEqual(ctx.exception.status, 503) + + def test_invalid_authority_token_remains_unauthorized(self) -> None: + error = HTTPError("http://127.0.0.1:8090/license/verify", 401, "invalid", {}, io.BytesIO()) + with patch("nano_server.urlopen", side_effect=error): + with self.assertRaises(ValidationError) as ctx: + license_verify({"token": "invalid"}) + self.assertEqual(ctx.exception.status, 401) + + def test_verified_local_authority_response_is_returned(self) -> None: + response = _Response(b'{"tier":"pro","issuedAt":1700000000000,"expiresAt":null}') + with patch("nano_server.urlopen", return_value=response): + result = license_verify({"token": "opaque-token"}) + self.assertEqual(result["tier"], "pro") + + def test_non_loopback_verifier_configuration_is_rejected(self) -> None: + with patch.dict(os.environ, {"SELECTPILOT_BILLING_VERIFY_URL": "https://example.com/license/verify"}): + with self.assertRaises(ValidationError) as ctx: + license_verify({"token": "opaque-token"}) + self.assertEqual(ctx.exception.code, "invalid_entitlement_verifier") + + def test_malformed_authority_response_is_rejected(self) -> None: + response = _Response(b'{"tier":"pro","issuedAt":"not-a-timestamp"}') + with patch("nano_server.urlopen", return_value=response): + with self.assertRaises(ValidationError) as ctx: + license_verify({"token": "opaque-token"}) + self.assertEqual(ctx.exception.code, "invalid_entitlement_response") + + +if __name__ == "__main__": + unittest.main() From 8adf9f7577cb0237f91fe4cfad741f1b4d38f810 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:30:36 +0200 Subject: [PATCH 23/66] feat(release): add fail-closed store package Add SOD-836 listing, privacy, asset validation, and deterministic package tooling. Store packaging remains blocked until SOD-837 configures production entitlement signatures.\n\nVerification: pnpm validate:store; node --test tests/panel/store-package.test.mjs Refs: SOD-836, SOD-837 --- .github/workflows/ci.yml | 3 + .gitignore | 1 + README.md | 11 ++ assets/marketing/README.md | 6 + assets/marketing/selectpilot-marquee.png | Bin 0 -> 84315 bytes assets/marketing/selectpilot-marquee.svg | 39 +++++++ assets/marketing/selectpilot-small-promo.png | Bin 0 -> 26747 bytes assets/marketing/selectpilot-small-promo.svg | 28 +++++ docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md | 67 +++++++++++ docs/CHROME_WEB_STORE_SUBMISSION.md | 109 ++++++++++++++++++ docs/PRIVACY_POLICY.md | 68 ++++++++++++ package.json | 2 + scripts/package-chrome-store.mjs | 111 +++++++++++++++++++ scripts/validate-store-assets.mjs | 45 ++++++++ tests/panel/store-package.test.mjs | 35 ++++++ 15 files changed, 525 insertions(+) create mode 100644 assets/marketing/selectpilot-marquee.png create mode 100644 assets/marketing/selectpilot-marquee.svg create mode 100644 assets/marketing/selectpilot-small-promo.png create mode 100644 assets/marketing/selectpilot-small-promo.svg create mode 100644 docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md create mode 100644 docs/CHROME_WEB_STORE_SUBMISSION.md create mode 100644 docs/PRIVACY_POLICY.md create mode 100644 scripts/package-chrome-store.mjs create mode 100644 scripts/validate-store-assets.mjs create mode 100644 tests/panel/store-package.test.mjs diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index c4226ab..e01de27 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -64,6 +64,9 @@ jobs: - name: Build run: pnpm build + - name: Validate Chrome Web Store assets + run: pnpm validate:store + - name: Smoke direct server entrypoint run: python3 server/nano_server.py --help diff --git a/.gitignore b/.gitignore index 91f43eb..2da88e8 100644 --- a/.gitignore +++ b/.gitignore @@ -6,5 +6,6 @@ npm-debug.log* test-results/ playwright-report/ reports/code_admission_evidence.json +dist/ /usr/local/var/log/chromeai/ /usr/local/var/run/chromeai/ diff --git a/README.md b/README.md index 4e2e8de..78451ea 100644 --- a/README.md +++ b/README.md @@ -202,6 +202,17 @@ Current CI enforces: - typecheck (`pnpm typecheck`) - build (`pnpm build`) - baseline tests (`pnpm test`) +- Chrome Web Store asset dimensions (`pnpm validate:store`) + +## Chrome Web Store Release + +Store listing copy, privacy language, and the release checklist live in: + +- [`docs/CHROME_WEB_STORE_SUBMISSION.md`](docs/CHROME_WEB_STORE_SUBMISSION.md) +- [`docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md`](docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md) +- [`docs/PRIVACY_POLICY.md`](docs/PRIVACY_POLICY.md) + +`pnpm package:store` builds and audits the upload ZIP. It intentionally fails closed until production entitlement signature verification is configured; no unsigned commercial package is release-eligible. ## Compute Distribution diff --git a/assets/marketing/README.md b/assets/marketing/README.md index d62df50..3eb058e 100644 --- a/assets/marketing/README.md +++ b/assets/marketing/README.md @@ -6,6 +6,8 @@ Primary launch visuals: - `selectpilot-screenshot-extract.svg` - `selectpilot-screenshot-runtime.svg` - `selectpilot-screenshot-privacy.svg` +- `selectpilot-small-promo.svg` / `.png` (`440x280`) +- `selectpilot-marquee.svg` / `.png` (`1400x560`) Recommended use: @@ -13,6 +15,10 @@ Recommended use: - `screenshot-extract`: primary product screenshot - `screenshot-runtime`: onboarding/runtime proof screenshot - `screenshot-privacy`: privacy boundary screenshot +- `small-promo`: Chrome Web Store small promotional tile +- `marquee`: Chrome Web Store marquee promotional tile + +Run `pnpm validate:store` to verify every required raster dimension. Store graphics are listing assets and are not included in the extension upload ZIP. These visuals follow the same MUE direction as the product: diff --git a/assets/marketing/selectpilot-marquee.png b/assets/marketing/selectpilot-marquee.png new file mode 100644 index 0000000000000000000000000000000000000000..7df04c40498ef755d8299298aafdb4fc902328aa GIT binary patch literal 84315 zcmeGDbyr-?^F4r0Aduh$*FbQB1$TFcKyY_=*FXXUcXtmVFvtMGgS*V&9^Bpio_s%f ze)k>RzwRv7V%E^~>8`HcwQE;RxQdcA8Zr?w2n0ftm67-e0wH*TKyV*k!vUYv_9PvH zK$sv|2~l;gtb-+ZKN4MH#FKFQ237lodi(Wqe_P-e@Ka=MN*w8TY%zc6;c4NSXqTXP z{&4){V`(`$9K*gXXgLbSh>kD?X8gYwhiLALe{aVzUlRR$D`JXi_3texS~Ty!5gdp? zi2n^H`2P?7|G71u%0#}XP~k-;Qsq(&v~OI4;Qx0Tal3AA1p4}<`lLwu1J4A4`UQpH z(xC^28o#`tg4f}2)gp9QuivQtJGu&~IgNNa+3x;r1NsdkqdeR1V9Pw(aPbe?+8Ssa zuS42{QN-f4_!7{=@`Kvia?laJ(h*6(NRY9j zK}9Gw9eZTM!zu1mhy?vf#r^#wcy{(J2W-&!@>QJ`QpB2tX_(_FL=H5ODF}#((g^l4 z33iFiM*j1}Ig;eUKI+%&SRtq3aMtGj?-jWV4;(ax_AH7;;3a+5)*SChZRxn7!@vR zZgd&`*}(5Bg|n8rpF+!LrNQ@7Lv0% zhubA}%r6qC{=0JhTXePxgl06n?Iej=hSN&(vk4LjJoN9GRq2vA2LD*u+t@_SQ2$%+ zcdF%n5kqb|0j_tdbXbv03aIW7@BpIyPv~zcLN|AJy@TO5=KsdVv9WVGRdbdyvYTfW znqU(i7P<$8T5%2L6()AYz;W*A|Aimm^@l|&)JOe;GNEO@3<-9rdo|)AG6!H{Y-~x6 zKUStcz@<|aR{zGD;lT^)5AsZih!uRRfkkma3x-=n=whN%Ar`Af_BM9kqobw&H!H{r zX^^K@kXA*5PnFID5?cR~NMfyMXZ9sg6&pL7cr)t1VL#F`2~=n@5KT<5B@3ve_^ssg z-o2-JM=y&NTV4IZ;@>C{w50HopB(1j>q@H;D_(PSqkIKCG^njIF`Dq-WdYd~{v8lP z;Vf35(IOIm3wTO%caJPlrBfB#6tw9`_}{TWh|G%*nW&>VyL(U|tC!)!7=L%0kdUAd zjfx6Q|7U8EK|;bk)q0(*q#`sy6*HH-per%GGXs%{{}x`>Yf)atn9DY=P%^k7BCh8W zm|{VLhDDk6%SxWAyfB{0MP`3EeRZ6>M5f{I1O%;Fk5J-w@x9Ocp0*rfkI!(^n> zT2^C+q~BgilpqjEP*n5K%W7)A%2xb0!0Ky!P)<&zQi2vt>qVk08`C&BrcC|Zq|E<| z4`{5iqgE&gb05aZ6I!^9P0W7(``6E*RfP8S)xiq!yg{KwRPz?HPCYL<|BEFYVIx** zC7CtL=@hPTXm;+1rsDqkq@@468pQFHZC+|1`dy9ov4YkM8uo7q~c{qO&XN;35-bG4&6 z%rv4yH$=X?g+u#)e@T6K68Tt<%uzMPJ&r~7zmGrnRx{z6X@IfQ)eGP8!t%gjT>oze z3M|BJd>sx<|KBOD%US#FaQ|DzXSdFJ^|)4p?0--A9EkP*hUT%p=J@a9Uh?rcIp7@W za8P<=#qwp;(Cl`A*Z%(QKCHb7@y_R;&xmMz5fMl3rbRYg$1+_J1p#}JwRO!Cuat*EG~|JS$gJn=l@nFmpY-=tc|se+*s6LoCd^>rLjSqC1R z|1Cho_T6|MTnK>}rD+fHJ47I0!21i{5fJU{+R!~KzloLzk|9s>GYeFu!U~h>{#evC zJvcnynfA{M$2DGaoD?j!IjT@?J`|f?T)HVLqK+g8a|{>JB6D_kQ`e9obBq(j^U%{V z^0D!a_Hau3qrx0#uj{0*;0ggtLKT?AQFuy8*Dg7WK19?r`x7B1 zGp5`Ru9u-wC|@ZeK7aW4L>)?hVxoSRPKxy+RYZ1d@v$tfc}(*-Hnz#@VPkoQ-hkq| zVvK)qlE;Z@bwf!b#L6QON?Kpd;p|#JJu%hvSXf#r&mdLa7f_5F6Rm9a-(KJivTvwW z$uCeM%)&+#qq7belc=uFF$Dat&bs&iKeZU-B;zkeIG z5xV~3NoCP3@SeC+in;Q$wz_;Nrt0>Ss{0Qy&hC+&>32m=mj6A4d;x|OJ@~chI{!@q z*`-^(38dF^rAi{dh*q8wUNAPRzUjeU_-NVFt~1G770SBD=rIM;UgB#=30P{}?yrKX zO{GI&oqk8dr~3SLYx3lHwEov9@-jFz`v*rRU4NhPkdoBh-Amg|8}XGQfeI^9lA>CL zN=eQH1gd#Cn7F&YyZ1(4-(QgQE;l?rsEpNTMJ+Wo@a3bet$C*z&!kn{w%boAj$DT*Z=;HuoT<&Hx67ry}y zk?f(-Nh4wd))ChVWmmLoA*W$xWfuJn0-2fP5a7PP&|`U67y_vO zJ>57tPWV?bS$oDis&;{#IJ&q8c``X8V;eI|dz-x2hW^%fZ_qy#9LDGEC! zHNA3fYgSrXO=r-ZyKR3!g=XnpS#2s|9ds<4H_*L1PRI|5Q31`mZr7v`C6-HKUgj@o znUSe4N*B`{m~yjnazQ?@b8vHVjZ15jfy|6?2#U#r?MOPH(^*Ges;sAOeUZ`l5pH4~JG=?QFdIhHF}*q~x`Y z8OQvsX5|Vx5)zAHrP*rmEddDBf7LB3oByU7c2rrxq;Oz=0zm^kvmxit>!icNB6PlV z6_i0Z+4z94-U00mra(uv3Ojgf^-Vd6Bchv%+uO+D$65>O>(>d7a;l;1eI6g^CGuZ2 z?tICjsqk5BywZ}87yxek+NkVeB z?>VC2y5(lA1(sS5b(c$J{E;c7AH=Lv|5O?azOZecRxHWLz6D)i(+D9_(mgaUze@G=((=t||3P52k6VIA z1DM~z=i+qMM9$rvo$*=($+NVL5vW~Wkl5REyo#plKV<+vN}I(XYah4l$*7P--ZX5p z^4a~^L!ij-b^+nARYbCT+LB)c*uF|GxO}}EVg}~ zDXp!mE80(SG+1^K4ibR}{T60nHUQ9sUP-P%dM_(xddmu%2vkFjT(#v^6A`0$frMJ> zBpxiXW8xV8-L_=tC@6GxB?vKjiV#9r@c9!ue982fvvZ1 zYiTa_m(mB6j)eX0ZijM*+e@WKA$_7s*{u?}lCHGZM^I6ZmTL^M#w5db?^0=}Ff|`L zdJw2cN8PR<(20t&Ozh|OmR+>84G18?vl$GfOBl~55<&e7!o$-r^)GW2G9h-PRx)yO z5y}qf=@D?l3N=oiKP_n5ky+QUzse>gprK9T%{Q%o&QL|9M^@0%)~Vj<-X z(73A;ZOzCS40z^tu5WBiKNw}yD2mg|SZrKOk4gy81=4YPTp~jJghgXKL68b>4zmps zJcvruzJJqZ9~<)xb}9kZJL^!4>QIHGiUnk>@PyeHl<<#?ORQ8CEvLuaG_y3cu$`+V z=_KeUl3dybY2Y+`+^luMO76bCx;-!9!xUemL>h_Uy0WWX6;h!c4&#HDW!vj*m%i3q z!@xs$ik2&xAkb$!YCf)@Xtr7H%*LKg4x#_F0A5|nK@f;svO%JWbzWOOr<#76{4nwr zn?ZdglpHgB_C@Q@6`oo*?5|>~^9vaUEq-4X(~vjiM$*`H>Qm8zM4a}9H{(r9rTe-E zOBNya$Fr*NgnvA*Zj|P8Vc-`*9l(iz7k-@&-!iz+&-`3za~^^7m>nNCI==d}((3;J z54W`B(!|3LpmE#?K#w246k7dW?gb)r6u=u0(7CqMF^Ng=#JLr?En6?F7|GPdq7lOb z^YH0*X{r>JOC_zWZYuaO%=&QXlLDVB>sWfI)G|IqzDR?nWNRVGgC(V_ngy5y3ezirp+i>kR7y<4@ z$UB~iz`KXT)ph(AL88XSt0^gXyn0zETVrY^5JVjgoNx){-1dOSEpa0i88wr z8_ctQ&CZUlOQuf5V#?s$2{>OKZIw2Y7V^G>ah9T>QBhG1B9dBRG&gS=eT?y#^mNR< zCKH~p=E4AD1J?g2E1S?~9?vr{t0Ei*&kMj-o|t;GcirN5d539MpeEN=OYxtL8-8E!CXf#~KzJUS3`Rxq(U6 z>TM)r}oIrWIs z*APjCEUxE2@e^5U`%`5?K3;)y$fW0~KR4~@&pQe&H_IPw+MI9xEUVGRn0@>%-Eid+ zgvfGin60SSjz0T9Gw;^V#% z_~@0mAaLUo3oyfjzDqR7aa-E5kjegh$H6v@jwWuw*aA*~;|gqOwM-OJJT)H%)2(j& zA?My2ht`f{OioN~R$i+@oPt^xG+7_MzxE#q&1T{rAhUFKQ*e!YgImx-CxBv5j zg3V^Jh!@A|kG#_=%>C}-VA=t?IFOa;bh}s4K}q}SRoEHRDl9qkf{=YJa9s++BsK2J z3O(1)TX3`Y;pUcMJg%3l)p_6G`Q3W5-B$Pn{_aGbT6r-_$?op1^@cRyq?CXP;iXQ3 z_G?ymZKpBb+iA3Xk_2Hv|r5M=k!c!~XsOvHUpT z7FWE%dLtfWsAdZ9swNKPZIc!N?J+SEX;7)}lJEcU9U2F>)C{SHd=L)U3(jdXsFiZp z<#PO8^d;&RmwH%1F_Ysb*7>CPUAG~ek}$s?jtF2=X)ViBD|ppU}?%bi2C929)Z^PSz}BJk#3 zdFUIMA46*%7r2?0Y8>MEZlBNb!Q58UQi0kZ4{K0bX$}yp zFO;Tfyy|Yg&L!kjsfpHovJuLY$$v;4JMeWmUYt6ACEW7M*=R-w2pw}TvNy1X4O>#4vW2`SW0f&>Cw@8$&XwVOG3uMJeAwww z7$oI!`{w>U1=yaAUhIn@-I;w>UBX%`8iAQmbpyle}T{j z89g>XkZC0P)DIODhuj-_mtrBsfr%9(her4+DD=YRZZSD#>6Nh81tSGp1gSXuMY0r( zN~$@TQPu1)qEQY@h4m&xe#UZC7szjB{3OdqGuqesl!bfDG8>Ll%blH^#(d`N?4SG^ zF%I>W=X~clY`N?#CqLL<8G=n!6jr!fx&KygFe@J317TukUs``8MKse130dXJd4;&G_rbL|@Gp0PZoCR;zxr^dzt8Fc@&SP@bswh(L3z80}r#I}|hL zdGgi76xjs}3k$5*OwDhxnibB)b%rL3N1tTKjN*UcS$Ti86y_3Ms`$3}*Vk&wSLvCL zQxb=n4P*q-Sg&8^3I*Oft{qI1ZZITpSZnq1*;fQD14P+j_(GO)Geveal$8>(-+n8U zshJR9=XQYJ*f}Xfc3k!XRFnc}W{qVgDF}>hQ_^(W9D! zLblEnkJ|JPRd0a0!oY@Yex|SJ99c4dN>qz+{?Jg*%8oeN9wvBx*tFdxu#}djv2VI2 zHMX?vLI<)?S0IXT{`SlkRwEi@bn zv|OKrpXHEif)`i;Qcr}A5*cS}RrMTqs*!@X_iz-O+cpS~X4twxtoU=OO3+enx151- z9y4F%I~e@@)Mdyds$Qv;vI?zJW4QZ}AqL09}EF zg+4yOlL-ZQJodI0(vH$!X;(Sz&gU%6^YYFIK5j8_%Z{LOB`%Ul^qdW1`8`pzJ2j#+ zCY@ZO46m-;Y%NL8#aCAH5>yyFSTq@p7G2qAuvv@0!)+XK^K{*m&tebrQ>GC1_xmfa z^-DOPhRe!E#gRE<_dO?a}0C& zoQq6??=f=6+h2%^D%kY&_RuT;IVSsEq@ky$XIA7_jHOF#E5#47UD`NZswG|Ew1hEI zP{=t%)p0t`SWRo3C5m-X{LK>uPQw%MB2@tWQ-pF@feM`8L;IRgCh$8HRMyv~Wo+zl zwbDITv)=)s?;H!NW-9b)xF~B!@e=avS8aq+JY<+RU2r6<)ZqdUbF0e~SyNQ0OP%j;L&^X9cmfOae zqB1(%VV86Io8>soazT1?5aXK7Uof4yt#hOf6(J)thFsHq`CbPDdR`+1t z!c#`IhyiuvR3{Sx=;k_43Lf+svD#U;g)g4~>I=7MofOYtMfVP;C7CO@%It!@mOvKE)7l%=~f{>oz-n0>-m#Z5z4{iQjEA>7+n?Vw5wI@yzu z63ctns-=@AEhTxbXado?uOG1(Gkmw;NR8HDHPZ}`h{Jh~bEh$5TT^Vjw?~g%*T3@< zU0P`q2ux~RZ9>Qv#|2W6yv=EFplsD@t>Feu*kA=DH=31SH4Cbia zj$zZ2lV4TMZ~nB^1R;K`mOAe_FX6COUx0QPFyaY&U)gjdxX(B&#w$(~+7f&G?h>?| zQL6Ab=BLTt*vD5X7j5Qc@71K!gmo-mJ@&CgraC5Nwra2YNJ)jlz;BNNFY#_S_pe75 zuIlWUB&uJ94U1*YmH9{dzfzXz^YVT+@A?G6(WBr>rhD#*trN{6L3n%jxiD|KfWKwBqHuJkjv7bx25QTRoL^Bk*P%at)N9GW8*1~nisoo_?h?Np*?qnKGmaK&mZhQ zbW%=_P`eC8-D_BwY-J-k@u(5BA3$g}8#*m=<;;}m&G&}mHWOj415riBLj~@(kJbjp z{htj0gVFF7vH1ELTpaSv=s|pC|MCj{a0x2?~+|07ohA-}3Xt&EpUdy#Lo103!Tif_4$EjM?3b zbYMTs_A3~^h~eR#EDL4mm0}MS6SUm-o|W>cuV`^}WJo`Y)V%Lv2j;Q3BbpV?Yx?5V z^-s;^fsmh@ktOnmAu5e#0FL?dYY-DWcq5P-F6))X9;vh=4KM*2c!h5A;3kh{DO_Og zMkP1oMRWRI=`p}wOw@?9O#r7Q$1N~j_lKv9Lc1J^ipDeteU6ggCK(v%`|2R%BNX%o zT8)JI0N^22D~G+hf?qG6F4aBH#9OPdd&k+vw!G0pHao7~NWM56<6by9Xwgg9; zN~7sD8T}ZfRw4hh6QoGLcMYAKG@D7;%${#hNw3jKMwY&EbgODeBsMG)+f9_NFJBxc zVOHXAVAMRY>l}avOkY$-O&5-|xOlV=jzAh3mRjw5g6FVtY0Z=W8W#sC{9AY={;X~$ zAyHlm1IUc12>}{S$0fwdL6H&^2>0~8X!tY{Unwgd!Z=$X`_##ygaXs8R4RmvxMOXwK}8Gr5bAEcfpREPT|Yn^@nf; zdAf<&$%OFQSxQ#wB#EnHlvKidyLLTh?~&6&sQQNK0xTR0@9dWUuVfG?w?!YgP~mDT zHSzJ9p~IN9W0jS|$kN=#w%CF5Un;n;B;%N?l`-5W_reW=aQc5&VDI*#?{io zcAxS}T%3xjDJ89R*kYsLq+O#L71ra$>GdVbYZQ#wSosfZT#R}j44IWZr5KLVPCmku zk&Fxu4X=!Qdo;SOJjt1x7hl8B3XsJUadsRhI;6Zuj zj3ZHHpsC4OOnY$)YpcR{2%VjjNq!?b?SHFTa~e)skYa5qH=0*K;^p=Q89maNoMC7> z2+=z3fNV$MQxnp+2OwAfjK%7Vbl22m_4oz^!CT66bKkW{?iWc(u7F4LY3Fysis)9r zqg{G=A%_6WDhHIdFHRsJ(7d9uGK-Mt@4o||hj;cL-#_3#!UrMdeh0Y`{^7a5KDk&uOMT*BXr| zs6P;ih-3OKdaTnvT83B-`rIn{_W={0<9SXt`BiYGeh9f zpw37KpBw3$e*86=!CDcOzW(3QDC1L81F*W$+Qx&Y)ZWF(IMQ!1ibgd*chU?Q94-D| zjod8fU^%3wIzOl~23_(xgo3q)C~4-klZ#8-5Q~RwkC=ERxm*6j+2k(YP7Xa4lm%K`}PyGw#nUriL};rEZ(@VlnvlkGNROK z0Aqm@E0{GVM!}|L!1y=wi<5*Ls=&(3#m@yIbP?ZTG!tt47c4AFN)+Vue!1yi?E_w# zmQ)4#a15(1PI$Sh>TK7S0zsO_qxoBF7O8)Ok5Bf7F!~ATErX!h%*)s8cxaS}8nbec z5MTj)8b(&XTXkK2@rc^U1nOmxNMRh)KK`IWYhH!LZ*5sHWlphZ|HMQa zI>cu4gwBswoeZppW@HdB80H=5^(pJO_#ZnrYeX|7F$@5O3gMSTR6mEKyC|qX!2|qE z7@k7Fb@qg_f3enpV6#V~Z`C16rB4qMyu2XBEnWBfkh3hmbWliQvlPwEK#1VuYLGa267V`0A) z|A&ul3eH65KzeuAU>4r#Hv~2)_7gK%tHpN&>$RJjS*bKzm=oRZ9fmF4$I4nT8fZyt zZqJ=__bC~S&rBqaNaphSHp=9;)N@TA%oMek3krB${Iu?xKVT^JJQkK+_W<^ZzV-OE|Z-bG6#w2a~Uko4WkwNlRsRx7L)H|b8hcnol(1bcoeDloYP$X zuk1KKf3|b#`hR7|t>IO@N`>$(#eII>`Sap@;S@*$Tt1KYCZUcu1b7SxL5c?d%S5 zIHcy}INWn7S{aRXVB=hV*C-ayB@;~<K$8N+#redh|$Z9)x%qfCD10 zB_BVDtd0+=Z4EK3Dl+a#`_=t9&We`bwXGYdFWg-oDUi`@(p`|_`^$v<9=Fnw^_ZO- z4c5r{3zGd#D{are{0`Ci^~%Jo*(VH{q_nlQr6eSdgx!8Pj-l7=%HuG2tzAgI|60m+ zHD=g0K>*YS)UZvdg+=KD?`|hPw`QX~GGK*wby3brRwL0Vi3WwH&HnoJ%PVmuvH9Zk!sN&i8n|_g!gpRwY|;qwAZiz>s*|JE(?E{?VEW%Y*ywd zQ4cSWOA)qQ-TNxxLx&A>YTV#PC@3gM-!OW?6rm{YjBILJcyseLZC1-VFJHTBzE4S? z$Nxw!j7t#^y26+AU0eO1!{M+!kA$I9Q%nYs@zyZqi5h)td@Le9|D%Z6S)OVwg>v;G zCm|YptJ0-k?m*j9Kk93aho0#{bt=3tY1zA+yy2Dx!?g0+y0)D9i;;IY;ej_{K#h40 zJdM{|pR!~9cUx?V)tv}?lRrAq7oPQBue2#9@}kTf2)pPwqA z!o^Pdbkp#GTmCg#X^_pQ!R?VlQTUXUgd4cc0{X5ot`1B>Tt!txqN@6ugi$LatMOvf zp=jM}Qa%q2=jvpJloFAyOI-|rAQ}9Scq5l}btp%P%&e7`E+s8Le|xmlnjIgX@0K7E zM@OoJ^P1{eZ9QAt)wMQ9f4-_q({Zuz&>UDSa#KUnwvnk4nFo)M>dmF4jFp6433r}2 z118p+`%CF>9E0S|Il!(!>A^LRmAu(=u|PxU0ae5shyygPB8t3R-DM*7`mIp(w{LD9 zkcJxdD*<*p6|jn?&@HG-OmaF&}Y=IL}(;GnWE6p+#Z7q0Q@y^{%B<7re*f?OGQYr_Et*~CLt*q z0Uncr!(IH&-p(%`tqh_?WS~92VB-fd7PYL1U#I5nZbA_!i`L2OqiZ;u`UhSJ`ZDAZXj#D za&xa#tK)!=%Wm;{6r$saBCXUfG+AB|1DDzNxnWUOj9s^rG3oVhS!9)ya=DJ;}CN zv~&4W?`IsBlR|q>E@KfvdDKpcYZf+kJ2dC8uOf6;Xi+z3lxe}}r$|KTM(5%3nB!yq zPj3!@MuKQIf}Q=QX01}&q}Bd(V)$N`K+`M z(_T7EocNUVtYJaQ$cq#n?Ci=-BH(^F(&y+Xq7#^}EtsT#P^iYl7o5~BS#jmV3jldgSDbsh2 zc%#yrS%BKxE49Q|Ul8x&`UV?&9eqHbo$hMvQ?(!cKs&CSty6PAv9bEPys5pt_KzZ} z8<(RsG!KOarHTmLl#9XB+sITcM&^%nfqfn$lZNCVk&Q!4Z#7X2{@=mperjkiNUe)EX-hL)NNIfSD6K>dzX zmw}zZb^YR~|Li24IMq>>FUf+qGs&Ov2-aj!ch)nhH~Pi$xSo{9($=2U(@&TN9wFpw zZlep>WVaJxiovENy4Xql+$R|g?NgZJQ)X+9OXg!U2b2E6%*615vO%n&pOt!f@Rw}P zL^0cXveE~v&2(AI!9a#u^pvcyu@%Rq!0FCKx5YfAt9l)`!0`iT&E5H>?N>?$ zBb!HJ_d!s&1&o~L=L`M#hsHZMo1zx;8`Q!%+ptkSzUv<;EYo%Fd(AJK-1dw!ART*$?v~KD^Jc$)N;&h)b`ywCP*2aH!iTxeMxOEI!def zM6*bBy7!eWskot__vr+r!O4a%J7}PBxni&n$Zj6fEPopLEbzkSeAx;RuK#S z6WK5p^96rL)O!pnUpab}^lS_vFn{o=cPt9HYnmGTV6DQFOY2WC{#AK}%3UQAf3V_{ zDx%Wd-mk!`1e{DwW?e1$0)VkI`Y`|1Norn^l9B?@2*lk;kvvh(J9Z0uV};ybFGcl$ zW;g@PTyi7W-`2bE>KDnp4( z2TZ&-_5f4NSIQeH7@1dm6i#FO8W)@D zD{bw}e%QkCN%KO2yCTVdnJ?dg~K2WU+<;d5TUo z4ZDj!l!63~4q`|3>M4cZ#cRZlEFy+{CFAp7@4`R+D)FL+F#_+zZs}K3P@B->curyJ z+~9e0jGlnsmQ){J9birIu45s8$Ehlv#)NjC^O9VCm#Zy|6_I=uiOp!zjG6bUJS>Gd z;1z$2+}7Pg?BAohg(d@coPz_1X%2#3-iJr*^)U@$<*p6WmqotHR9F*}_{VMIgA+4Z zqv(S3oKEh}Dz}AMuZ9+AogOQ(A!Ec-jGhM%_Iw7d{9Me2HF;--s*Mfw4(?~L?0e-E z7!``^guq)~;5x|a61^pCbvhKfOO<8*QLL|1BeQ%pVpFzAtA=Nf8!jQZz%Agudz5VVE!WY<7SN6O_`)8*ybEV~R=X1TaUZGBi2`mEw#I}+o z00N;AQ1CcS5)*S-OWu>SrlsKXOvUX0il?yOWl16z*mCskA<_KY_sdC{(ut)$@NU6t z^ys|(0-cbxk^N8k4BrG7%)u-**Rz`dPJd$7g;+1ixiCmc|L}esggn{F?e>_PnE0(P zyEgsWQrvMgy)6AyZwVGA<@3*0LM2-Ezrf0#g=h|?(B=90-YUIqj9%j>EJ8|JWWb@H zQ=n-yyffW)gl>)315MD$%o(vq7!X?1^)*ZIY)_6CBy1(%;=S^9HR312;%n~dR$50E zy+8s2EDIAmb+8Yw;L}t1&PL1QiDt6{dx}G8(t4O%wSogqsAHXB`*YLv4K5J@v?05& zwP<=62*>qg0?O3srigB{9)AQ@(C|d4fj0Ce1mG(kI?9YnFwWn)qT0A{X!PU*x5R{1 zk`@0oYG1hZ*9T>JByH!@8u{zrY0msPzKUkg&d+~gKJ4+f?WAW&(`96TZ(L(czcKR+ zq;MvWm&DY5`)6}Od8HKfSgAUECrzv4csh-;&=d_Vphpfpr^(}Tl-&bQ2VGF1{qe?r zZM@P>iIbczDkt-rq*X%RluF-B-a|}6og*4J)K7X3U>&NY9{u{=xSX7~2ne&Rz%@hz zS~_N18#lHKOz8ZLoL=(pUbC^tSz^3KFZ_AAcs6h!E#DU61EyU<7s-K%WlFza_5)=Q&_ z?&ZnW3q_$n{D)OiEGCu>&tdfwElHerS6zW34TfNok4mv4l~um6Y{TrS$0Un|lGPho z1HFQ!bJtz;Y7(|G+G_JDNaD?28*U$;nzLR-2D3om5-wNV6a4+9;#wk;o+AcZTTS-k z!2^!T6Gv!BkzP+*6Z;s3_)0_~!lJ?>Pnh~@917{m?pH{@=s)KIvIH$>0|P61tL0%+ zd4D9m%!W%Hg=cf~%zR@FOM>=rsiPxQ1AL#u84hC1W9PS8ODQRAr^|G7M%F90e3)0< zWSZp2+Mo1V0)#_gZt9;t6=VxoZaVCCT90M5U!8<0kB=``oOx{-VPG@m`iFlhc<)v{ zZsF4;7xVgX^|Dz>7XNaB#p*By8z-_%PvZ>kgRe6*dFY5+>Lceh_&~a#bE}p<*^y;s zfHLgZ%!vJb@EF%6$&uzTk#NYPD(v%gic%F2_=No*tw(3u+`5u;ybfdu@KfVj$dZ;; z(l#_J>GryxNJ!(?w;k!N6}gy&-o4+Nafk|!v`hv2)kWg;+ne|;^a*?Qt)Am76uTGF z!5_jhlI3-DGtQTcUQ8Mr<>8ZN{sw`p1oLI=tn&1|DOAYd(8?r&K=zRcMGbz9E;RtC z$E$z#SbvhQFLIVxslRn?_I|uDtUtxXZTRTz+oiKv8nsh-gUniN>58tNA`>zYuUbv) znOi_eFa1uJM0)IkiAcRdz2rkOww(0GtDlOAM0M|z{Vt#OLzhZtO=RECox8aUIj>$7 z>neZWNY7ciJ+tHR%~evA!ySX7e0?=p#I$;=YiOkU>%fOjQe+O5 zj)VKL75a1G@6wTQ-iy|I!JNMIp=UQeuabihN54>Su362L$X@z2>Xp%b8Zlg5Sbz}E zMNymMFS+u43~=n2_wUGGRgY2#*?{=c_qy$`GQ6)mhW)7IyO;0~^o7{UwA1PSPzZC; z%AJO6^^GnSV8^bhIq;G^&(slC2%CSlR?KO+>qjIy6808<)v8*73`wAfPcy+(Pf^-j zfaFf`9_@i?&0oK|w%g3b~Rv(Kn*N;y*vglo^B}p*+(cH+v4NubepIVIx2>cvi^Ssi)g9pWkALEEA z4BEd%3=NwKt)98RtLDg!Mn5@}eT_KX6*u2<>$%)tgs4OP=;saQ5~UpqlX6jQmKn6( zbD!@Wmg1P_W#W!x&al%ib)?X&{QMi>J#qQkBJ4Tn+S{qAUhkZ!vNa(iy4fo6AHqiZ zaBNojgAX&OY*cCm7Nov#Bow&R-FsSd*$;3<918h9((^DTpugy0o9Cz;{oZ8pB8k1o zDf2b(6(Vw!yv$D?Lr#zK<5{TRhpw(OgX}m>2Q#7VQ-k=H=eS9$U{zVz`FSD!@omZp zk2hAb!SoNkkASlLHeQ!AL0z?^KtFz1T={9SGReH)BmbNLhM@6GxkhM>a?+GBgQg~f zVbmEaM~QxEI=7=q9}=WGc>5cqH1SC^ShLyB}GIW!iU`t3uAy zEkVus=^sFUNB z{jzNBnAE1ba0i*7TPQ3oFW&;69NJrRGXrF%)V)MfiD)|6gGCot824V{hj9fmPC8o& zyVSGt2}I!(?jVD^Of8&rqQ!tZA|BePURbc9cX@R?W!Z7_E}On#05K;aHlv8t<-LN% zWZ{z!tWl1Zsi$ghif)Ljo(P91-{rP~ShIABc)1j<`2=VJuSxdzdkEfH`7v-@tGF0DilZ($&$gQW5NY*`Rr~ z5wH90EKzjCH^j{!07Zm1JBAT0zELYLgw9hCg}0t6vhDxr!#+W#{Z z{u4QtF|V@f8=c2z|8fCxEttaZiZB*3p}@Wv%k$Ze2SVBv&;X-4gvaI1rcoA$(r87T zrZ~a{hiKbl>>X+S2lnF5R+7jutQs5ok1T zuVjxjs;q*9iN$VhxX3ySPziA9SNu*QTrS-h*VZ^y*i)v%YP4!hCvS1vS`~R8C~jR< zRRxkCbAEX(jPz7{tqLTSZIv5yuw7Yfo+ckp|B zas`t;og#(PFS+l5EunBZ2O#qnde$B`o-uJrH|?cg&udf7V2H!%fInn%J_x?iJUs&`?L4_iq9HuV~*@oR5A5l(MY&pY2ihunGB{ zmW*Bcj?tp-7L@kFr_(ihGd-%~^3otv%@H&*bIolrNe(n;=<}>Z+CZzF)$kmQhtlcs z53L*lb#tVCwN%vRDI%d&ut>997sFuL?_{miaNIK&bf?#L>Gkl3S>3Dr#`XPmc#;9p z-CaSg#azi3b@frIFsg&Hq=ms?4TVSQ@9jA@IbklN2L4WGYos)ddqEreNUw-J?6p(2 zN|K(7LmWTZ;beywh?K?GcRjnqSoPJNb*2n&q+a2m=B7lDxQL$wHS_(W%aY<9}=JT~9 zt5@aYGe~t5H0F4J*e*t^d>PTRD;_BD{lI_bjcjdv16kXKK&1K+cJv9_Ym4*4!snSF z`)?bEv@G@3NW3k5g4r|Y2UGVKjfryZBaJn&0ux;?z z@!;*ZI0*m!wp`%t6-r?W6{lvV>=EiFpg-Bcr{FsZ04ZJ^-(x@MyP1Uv0dZ~`9VQXc zA6`rlL^W81PDzeU-WZb%f^0JqgnU6I);L!dSzjagP+f7sn&9z=N|1EX|4eE=;wfZ5 zwuEIa;LpaVD+b+ug`Op(c3*g=9D&H#z>~2J&!O}U^aZPwko_y$AKNh;e6E{M4^zIe zS;;Nv_BJMZ~_;q(SU2;WIo46#UwLJlSaIX zR3i%9yu1fPxHaqrX)uv5FWi8%E{>gdrn|liEVC&DeeF?5Als{);9ErLz6TR*n@C!BKN)1FT`Q0Rt0{5=g z=uoI>2QIiXZ4%ah^nx+5f4s1uh;Uha>rRO7C#?*Vl^sF{sFx2O2{Pb5hHE>8vLMK^<5dBy=@x z5zo5g3kQOzt1G}Ql@I{1Gl1xzp2cH!hD^h|LQOiC?`^-!KY(VN(ti1edeiqgvy~ZK zYAX^~WQ$0&0+v_f3W8(@$FpmwU70$KrG%V&Z@GBhZXI&__EbJJ>KB@>&A#u*)}5Y8 z2sO~tKTy_hlIMLf?}twdw08$`{xPJ><4zfDu+EXjC4C{^TbXm0k3V93`TicITeS+N@j{2| z%&~B3HqFSL93c@=5?$8y+`cjnnp;gnJeM)Kel?|~gMimvN`m*}Nnu4{ch>Qb@3NE& zH&d;BDy1k)POn@@VFdF3l<_IBzhV4|6%Vf!=jAZ_Foi5X_YQPK~g$~2I=l@>F#bs zK)Or18>FQLM0)7%?jE}Si|2XYeH?p#+Q$Uy@HXRXk(DoPzZtQjhhE zo-byLJ`XaoE3o7Hw(`N;=+~+^uiQzDt8GUKceWo6icBMDy^K5}b66~K{0ENeCpQkM zIhkS&V|S?My}a|TJ-EA>W|Vx|5~1XoqIZW@1jB)H9J@JNM^tkn>Nh9;T#`%f+tbZ# zM(>i=zy1Buqn5n!c=K#FwNBbR370@2!FP~dezL@_3<(9N!7XL7Y<%;w*%Dd?rV(oh zP6_$MGxgdVRl=qPBTBO?EW=e_l#M15?V zjejQN%Mwh)t~#=$=PMkhd0WUuCSYdiVT_k5#gv79`9kyG zwir`;lXh1%dZvibbZ^>Oe1~G1jU=x0qG@O;qdCrzWvp6Tkenmw>9mTU$xc2;@@Wed zn?!g(Dh&A94OZnUz$x6rFx)8TzJu`VV2W^`(e+eg- zksKUdRQc6bEsS5`iwIB-sp_IvOg`rqc@y#%_BL2(m5jU~2a9_9Uh)ZC)YDn6XH)^h&amFA9^0b)8AX8nEC-NuJw|W*v&aGy}~#e zeGJ0Q)>kO%~|01&G&h1 z?|>hwA-b`V)Bj;=VPHJY&k*=*FNisd*jp`kvYpJ84Iboo>L=t!Dm`O?LU8}UDj6j) ziBhe_yNRD9j^JopZ>q$@Fu#zz;4kWY6Pr|9=6V&OdGdW-s zLDkeqy4;)jkXW#HmUzT=#7@b@b${3=tmt_UzeT5^TB+4rPNX^9GyVH&;a0ZU^FCc# zu0^7Ojn=fwaeHqUwQ0%cj>%rEg1gsdi&nj1W9~AeA4g_#Y7AVH{I?{-Bwvlq&GU=T zISDK*peG3l*lcVnPu`E$@bLbPwU0B{bZ|_6ati1HRlvf=M6=rb+5K4#i3otn@98@2l$MNF~}?0ty&m8a73^Z>_+R4Nh*( zy~VT+1(DA7H9!-_{qoh$7O-RX#ZeoZCTWc^H}DFN9+_hi)a~D%4$iR_IJA2<0LhZu z{Up4^e4#h$4!!k!qt9eqG_YfEPW%Pe0_uNRa3lmL=g7G}7D}|LCGK7JMj?uuOFcbu zoVDf~RYsCn$|f9}#`y=DkQgF|tp~*rrZ_6~8r6C>*u-iJgOZ(wlnU-*7SPL5)M$_C;XHlO>A1l#Z%3<8(!=>|HYpWr6STzF&J zqf;o<(G6K{x<0nC&K92= zAeB25Xv7*8mZ)z^rIT@XX2Wm5h7;>u4WQF)a#;TxPy#vJ8roBD63I*DNX=CQVs7Wl zc13*3I6m2yS>?FT6rAMF8#ihDwvkXUZyogVyvGx7Dr;NCtTDV{*uDBL`+ue;kgJNNJn7iE@u;L26W-orhnbv^-0mhxU5RfNqS3iCR@4n{m--Q|suTR(Ud2`WGzXyqa)@l8^kgr{( zs7jkQ(I6;^e|0nHvC|hG(l(kae;{&hQQZm98V55zj&tSy?d@6iy_C`|jT7cmg{zyp z9j2IAVX53+J9l||8Nj}Kdd|f&*7NDu(Zy<{)H~D|6$st_=FXob+R*E?+THAL0O>fj zOkRM?(P*7RcLL_9-l)2J1!5O{`BJ0H#_Yk^RypPSFA5|+>A=8VXFWnGYmE;5BxfpK zN5BVdrR|*-cSuXy&(kF$+Rv9E>^;sA^xT8`hpJiwV`;iWS6>3%Vu4A4aroZPBwS`H zq3dH{Rn6db*f}0ywx>YH`WsK9S?m&dfx0&_Zdq9Yr&?agjvAC~44ZMg-E;`~dB$=Ihhh=}Z_Yk$5fmKGd8bR5XEk1XrkyViijq9uL$nc`-4DYsCvJDh>}E+_GP(i($xKrml(HAN8F#5%YO@ z9ACZkLV)dP)!6bnQwtlv42~(0GI9vRyfjoKk8SQJ!IUr_nUa;p^Y!yh_|JQrrnhiWtfcT(%jkan{K zF}_;0XT&Lm*}43~*>o>*XhK=e)wdW!okc#?(isWjK8m3fjtC}%F!#llM{cW&QJN`h zD&4CAxBi_mU^L?8;Y^@TrW2;zX(6wnxFqIewRo_HS~g?8MBAj-5hU)r8cb%CVZ6N@ z>yA7li*`8pM>JCrm!bO3^tb2vDy5}f<1gTERP^Vf?s zk4OL)+TbsdAZ`=^#!#&+Eh-T~sCIMtooQ7YT|IMuJXL$*hKv;E{=9?hWb=ju-B$q! z-i@4&qh8ygl?Ll(>Dn7Sye^kF+S(3b_>++~E0|Sy5HmEQ*InX{KPV!J%TLNSF|mlM z4TipUonZC30hi)?Zf|Hqo;1N!7C&HC*gm$eQFV0LX3+#>cfzuMbUqMSVA9tY3@sHp z(DvkkwQaeHN1-+1xfSyk-C7vE*hqYpaVOf1oLrGl22`0kUXbXOSXyL83pjl=x|oZ9 zo$Macv03t3ABjyIl7-mPbMbyN6h< z)S8TIPpK1C3Vn0WX1C~8lbvpo8J zh|R25Sim z9D;tcz8C0e9~|{d+z=IS7%`0FOJ>(o66;sT?F$dB|W{C!?m zS58x{YhXWUb=w*#&FRe8wFkarPL~q|FYCo?2juYZP!rKnijPbZcDcIzZ^DZ- zDp6k2AI|#-RQd{bLk~DyV?I6yMpzY-G9V0c@Mx5OVc1u?5fqt`!EQhmUaH)Zt;J94 z)2cN=op^ixl2+h3oGHx<{0A>!iC5-}r3$j|fDe>b+6vY(KlqmhKKXTY8Sp}+HJl5W zJjplMY~~=Y>txV>mFk!;@J^O$UP?|fQPmrEiBQ`cvXH?~uJ?|qPVCifdUStj5bmF> zKjM@-F(M3qjf7@4(~7Q%OrFPij%$DMEd-CMZ`k7xF8Qa%OFX-WQC5+D07@%%?!4Fy zn#p@BNb=kf4UoTHK1Wt&s(=LWo-?mee+C>GW6$ymL==+Mop7s};ojbF6u zB67(jA8!u2nQ_X;w)VUD{IqXUx*~FA#Y}&%iq&2CH8kuLX0;qxmYJQ}j`%YmVfr{6}{j--i7Q947K$Y&ZU;Kc>5I67r1>$`un9#{68lAE%<8nu@}5kEU}=V3E3X9ub_oeFkQIxv2>9Cb_|W`5uc89t zQCv0Y18SYIED`wJP~di9C=$p*-EhJK`vvd;nzimOl>+31-DapvdPC3kHy4R1+s*_Y zsoC7ipMSf$F2$9NHK7lJ6l1NG_Qp`n^aM)y5|cK0rP&JhbcsKKk;(sZkv0Z4@RNFk zV5#?QAJ&p;vS`gG>k{>OoiAXIpykM6lFZ=Ls&xcYn&q%~K!W~7HkO6i4)q$sJMzuA zh&kn;PzeP6KFX{#3ugND#Am7N_pNFBg)s9*7||#P<)^S@LMT&VjTJUPP}sQw7ovTB!KghD#f?_f)zwug0`OYB8I6qxF|c>; zF6LRa-?>>@Spmr5Qp;nrOuVet*77%X%HSs+A`N=%e^#yX^D6e_Tmxid4J| zjIbw~V~edTx`161Kp~OalrG|j$HK(}OI9C8qrFaci8^x>Z?S>xawr1=IIt@a`Xfp~?VJ9$2cWOz{3WssKr71S%d%@`*!m7< zXJ!jIJiyORe$Gf^?#@6V!|JBhXtkc{8}zIOE~?LYeu1C}Jlmi~Kh&Y0^2-aYUIRqA z5G<;{>Nq#+JCgQ(fq?0Ag7sd3&2n`*YgFY1P`HL|z>R7;^sq+yutxbFFxavTOhA#_ znJWQ~!yBF|rHdWMlnRuOt!Wm9=%4TybC1BbDU;e-n3S&Kw3XiJf8$c1m2LU>c8h*t zxipEvdtiCK!S!}gvq~RhDffxfwmqwYf0u1lq1Niy#3qeHr``n41zj)`sBRWkW})AH zPZ^gRDS56z)24%QSnR>Cs19X8QRxXwVT8UsN#aqGCtm{h?w9D_xW0x64>p#;S)za$V6 zSY}D&*`wQJWR(V3t{tOYI5#$$r@<-Fxfm`XG?W%aN3X2gbxxI~c0;Bv;}WrMU)qY| zfzAgtj#1R^e3nsY{yA2+qPUaPe#U^OrRDst(;R!^aJ^~)VB%3;ycFXq(A=8bQytt& zwdg7yEGj8UeP+?2O7mU3yvKI)N~S@um+Ug3(QP3c6ip7$TCb@uA{gmyuAQo;%})2L zi#(i^DwSDqGXq5iSNfb5kOwT^Os&e8kNSP*Wjbt{Y<+*of=Vuqv}jeAGj=ZCvqK>} z$E~PUfP#m9nc7re9);ZWi!#XbZv16KbR4NRw+&Ngz^EjFN!Brj$qSVxcA)dkHdl)alq*A2=Ke{zW) zgQnYPA6xzISv7j-feKj^3O}u5YwSTec54{%Er;<{KGo)V3DL;te+FgNfJ{G39-6uJ!ft2QNyUf9i`R z6&124L`cSr8V1ZmbDscXZ)l@eQ}gfiwCP~gOlmAP{T3znP+Gs&&H6~{&K9l4!P=hB zMTau<&l~Mi(3eZ=tpW(YxnDZ8P;tk=FaqWzn}0`(ykUX+>%j=4!V)VC4X=M-YByV` z3~XBH8UP%`liSkQnLl%731ST+0oE+?TI_*zJ-t73lbjneRLI>NJ7+lbFph{@ zJu5O$h{iRqW#!C#z{<=)(IPFl^hrUr!mcfg!JyA{gIPmeBg{>{nf-rt0mOr8_Z}Ck zvRHwjUOA!t%D+?|rtla}*kk$%Db2Qy@XMA1Bt3^zx!7<7dyswA8j~6c5mAwl@0YG> zEB^{uJ@Tct%Avn3M^HXpzBqt9w6#m*b8>OO4km&fMqN;Aw@*NcRACnd&?9HUcxsNc zDDQ`39Y#Sx3#euPxj<-r`8(Sx>^f(q@W%^P#NEDmx;0qH6|82>4~93r^yuTLKb(%N zPJAXXw2>!?6?+@^Szd^35ZItNw#roUp65t_ErmVwO-pM*zAJOf6xc-;k429TiEx!Q zSZ&1kvaa5~ejzn`ypa)#*>|+MUub_3vLWdZQSb+eiv&BAnA@`>*;ud5Is%Q_942?- zU2s8*OM86Gc=^^+tsSijmISvo7+g@4(Pm z%#DY%In>YbP*6dfgXoDvrESX7SZkoPGnPZTM26Zr|772b!Zt-YT}td93m2!QCQ+V= z@V9`Dt0qS2DxsMGHVPE_8B|EB6=o)VacnjkZT4MNU7=X3C48ln;5v|S#OfY<2bz%E zIrfMg4vmy`?H4BwA-ARjzN~Hl;;p4*X6agn@_^O3|b&mVn?>Y;d`&@-9bSR2`70HM+G?Grvn7VFPkr}+7LUB|G zWl4ijcWdowU79}F16}|wM-6doe9@SIl}52|(b_fgVQ7uc!mhp558JBh>MUEttxy48 zUT-7jCer^HlM(y^eco@(DXa9TX6t!Rpl^hP+r;J=`gg^Ix1ec8Xt*;-Akl6$yWUrxsSN{xXQ78N2-EFhlOt+UFmr5!MKO zK-E#>cym0sr&{FUDnzD>?~?wL2|Q8gy7u4Il6zj473v#?qVt86`cjX9o@YUkX&WAE zB9p~=)d~wGB?&z#bBadz6v-kEV}f$FueV%f(N#rw8r(X%b(UC{w4Xu(_^H9-?~@uU zNImCk2)1iFc#|09bv)Fa*`iv@IV6c@*K_N!$rkv)2ztc^yUHpjoX_grQ%Wf@pH1=e zVBy;+YO%{vw$K-cpwC?}<0TA*{N@Vcr_~JR3w}Ehj!Ablg|n08RH}Fb=%Y*&PF?(2QLG}xc!!Y; zizdS^p!7m075998DT_-$z?e+U$XGTgPEE1Ofen}!VY5=wQb_~5U8XS7s`^TWDkA@ zC&oklYk2GvXQMl3;7!l>hg;jQT5SPO_jo^xHh@(fp>Q^v2Ru`#ZnHO_F45O6@uzb- z*a1(oRSfZV+2_yfV{w!!rU0!iWe(`Rn6wHc!qtX}@=TtPf^u zt^0BlW{y;_KD%C&Rp|2TXa5%N93$y>#Ed2;gY_Y2UNOMi+05EYGq z41Zx&1FtMZjr>OdW=-yGFh3;|BO43bID}21o{iX$12lv4e~3-k?E5_Ib{X2_=aZxE zPUND;&THHPd`||qhtpBJVhAr~a;UvDXpi37tiy?n(v!Ggk55nk*1sN=Lg@2lwfj}S zLio5QiiC7A=8r1-Favl&q;uuDU81^tJ{$KjzdS+Ln)XkkDnPSH7mS^c&q%`LDK>g4 znV}ied65b~YAUPOc>nTEL?F1I!9M8Qm~_JBN_&1I;iis|AVNijS$>v2KYybI2RDcj z@)t1}jFzu{Y_roiV0uXkXq#PRIXcnI5|^A@OJoBU!vKB?`6u)M13Z{EIQ4tKSRHAEg?KZbI6Uc;L;8|ihv&K`UmAb~~b)%AFE-W(E(ed%0 zR8pDos>zV9&5~20O@$0f?w87s0l1q;Y`eM0o;G;A^SOb~X zs8?isL#1Eqdw94gYsw!*Hu|Q<`jl@%J{qj<7}!`PWb>|lxm~d}enCKR2&Q|^!sKsW zzvuS+2Zux5Zaq@dJ}4OkK%AY)%lF+w`*2u)R7w>5m$Y?i z+V2h+wYqo$I9o4Vu7qdyA3pZc`M<-HqnIG_q=&(=GiwfQ0Z5L)QWAi2CkVzquN%03 zhkiqU&ymOdSZ#3m)ymb3+dxQ2B)(|8P;I8sKEtn*99)A{hs^n5tErf8jtj0kspI?X z^OLy^zAD7*UX)QfNa)6sqK;-c~o`wqJ3$F?pTgt8l|w`ghWDgAsTrl5H&Rrgija)0AK_Vl}bhQFu` zO=KM|zWt@HDg79nM`XDb$Jo5bF>DWI5utawp^nWs|eD(oZYiZ|~XRxLx0zV7s8>4f{LXefwn5;#26#mG=Bu%|poC z;j$iuv|@|-%hs8K0WH|>phs&_9HMN&E{Z#M1%zfs2kcpZqUj)m4{=MEEBrXkXR(OX z1!&-@99ruN3VKF8Jyfa4$;@7*5fRYRA_MjU;G92)HT0vjDOoyA<`wjNeHcCr4~bYw zH8eM8va}_Jih~X)w<8Rt(`IXt)Mn96mEIhTsL*am0^)Kg1s%1X$5R4Eb#mWi9}6ej z`^rEc1VHU+U!bK}%8h=MUK+x($c*`ie?Dhy_|_tynT4fLIR|Mgcc(frMCfBHo^0wS zG{d8(=5cFVkMGI{Fx|f9QQ#c)r{}gWkyZ++B+!?Lm@NqD>vzALa z?M_a^zveVlrxDH7%UM1SErfGhtHMy|chG*4^dk55x?#&jF8Ap;@d;{{*NbM8ebBA#0DWG6Zm+!0sVm zdBXNYj{%sdN`~;Th;A;-M}!ynOakfieHtSK-1-WDP*s?F^Y7pmax}kG<79L6EC*19 zhpty)ZQ&I{H$5lSAlpG@0P_K{eY1=$+!DVorgKcP0US-|8j9@Xts{D`%(t`484nXl7eF?(XcJ(Asd^$tRB(>DNyk5BXx zhoP)Aca)WsJcl5a5Yn^F72Xn6(Ehkg=fJ9uAmWK~7zMzRJ(feUFK=emW z4NI3O`_?YNm?iIID6oT}e7g1=;~nUjF4mp-f<86ir0o3NL3Y3?H=nQKGJ+;MuI)ex z3<-IxRn*j8Y>}Bz1U>U=U=B|#pZ}~|+dHzhWq5>+sn71Vso#mcux-DbS|U2dW^Z8H zT|QS6%6#%2^e}^&^-cMp5)ZJwnoHdGMRU8KL!joZ$1}gh9r(N+{&56%V2}pdPjP>A zz%DhUAFWtwrfTJX9DQ4Wi;>poSuOXMHhi#FfDL4D2zDCC1F=j^NLsuZU`45w-UR3m z+q>e?lgj<8&^>};egPosjak|oh8dkiG}^?^x1pLxNioVN{d9E1W?mTPD2n{8hde&FOL?C?(?-JHXEJS(7cku1g3bc(6(+;Ieaz% zh7+qVv9h{0J8CBC8INQlHF=uENFiuAIMBr=Oxhf)kzc3XJ-ef=V;!hla^R9J5WD>shwL>Um#7`1(z!)<{DNT!1x65 zDy5bF)|o?OMGMr#aycLT0qrv7nezU=rE#C31XdcmS9BljpUh8YhednpX62X+LWsK@ z8V*H6w1LQ7WW~PxZdZxR<$G@B^m;H2ZL$pffCbiKLZW>a>k)ow<6A?hZsoCavU!IgC@b2EjC4pe3K``_kw{|iLF*7Y=^sR8;xRn#OGJ}uKC&eA= zd}Kempb{U^-;|lv7JEn{7;NS0%k8?MFMJg}Pc?`CLnw=qKP{trAI@rM2acF;`e10Q zMGO`5Jzu?&_j4_}RR-~=3`X+-*^jiL7(C=nFXtA=H6rt&_HQ(=98naT>0>r zE6WaIQ$beRK1dbAb{D<(Zn~sq3Aui$my6s3>EhV`}mBvo`idH$BGB!=F-1w|9z{|r)Umb&eshdfIrKHVxhEQ7@Rf`?eaG0;zee7L!tN4v1arl~D{9^(S0bU7vC}kj50;H_a zWz2wxiC3}I0I|QnkmDco0*uCvo+e@MFzaynj2U9lsFf)fsx`UDl{Hu_Z1XWw^9GKB>t2ALJ0Rx|z7_yyoe$;h%51hK(@{~seW>q4{j$nO#V zadD}kv{x<_2P3Ra+(m%ZHT8t&p2c$Dacg|JA?QZXr*1X(`|n?&8OZ{cczryoGHL& zD!dwiSOEzf*|UXDqLSWpAek{j9-THyhCp!)X71wEceO&r`J$>=waE4S?n`ST&Xp7) z7wmkjBxK>Y<#ZJQWKE@3gC9y*Dab|i1=C!I3xk-1GN`>qz2)C2?*a)h)8O>6mjwtY zMHZZZe}_6;=@&P;Q9SaQ&*-J$d=ql!l(e2VX7qHtF_lXUrgB-F=;%n5Cy3{ZBw~$X z>a@~e+w4ytxvI=TSZK1V0;EXTL>$GCrTM99zF1^|;oSUun7n^5;$>e_auI72b~ z)D91+V9hGzzPN}tP=b}6N{3)e?UQlN`$_v7Ei|_u&+qe{hWLp-D!CW zby#|{!u2EHj4QdUtZe(ua6m=XS2P(+qu=(2aD$75-c}n_W|Bc@`{D&-fbc_$S?*P- z)@*;E0>!(%*v<7^pk(-qLk1!f7H`oB%RKV!N;vTF?5C8NAc<}0{RI=+UJMX9Dqmpo1II}IZYg%@9|H;knYQu0SGUH<>Z)7q*fOa z7kx(^5G3mA^|-UPb+m@@1|M6*O(RRp(9BwmcgkR6EpP3k6cp{E8~No%HuofawKm+F_3^QVDfr_lTuPjba?XLzxKt5 z)we)oB;jO2xkGWFnO1vs^K#hu=IUyUEmSC5j#Db1Tty`zPXe<{t1_%_#>}O~^L}w^ zh>wJXorK{>c>uVlJ-E0$xKi$}9Tdk!kkJ`)Gh`RrVK(hr5n-EvNfRx9GKq~}*m-_) z@Je?H?i1qa=E*>u#ofx9$;yiS=4L)%UKufAExah#ie-5rIpm6YX;Sl#*?syo8Z#4H zu{>*RN%{!$V)h0B;DT*Q5vKst&ao^~WJq-t1KaSFp&slW(mSTddx_%pd_YL&A;!AP zs%*4mvHabC)!U{yFpw8X@fy_AR!bWVD(U+Kj4$i9MK+>~|#~ zst=?sgtOxN0lIT15H&wF-u4B^WJU)I-5K!|pB0F-)YanGWH31K#-b8}CA9dFpx%Z*Pr7rQkkk0(u+#I)&cer`*3P>zr4m^!@u~w9q zAeF0NNXPADH(ZXATBp=@DuVWUmip%KpxR<78lAWmINjoCwX^lF6+pOGoV4x9sq{@e z&=e-6Br8{FWC1Alyo>LOXAj4ZikbJy-OCYR;}J$Wtu84bXu*(2=4&tYU8zt zR607+Kv-pMtxCfH=z9@)pz(O0#yE@R+^}laFI#a@VewTO46V=uI3GWtAD(Rtc0J5z z8|{Jgyu7^Rrj)&Si~5Lk4 zG00=9KjMWwtY~I=?f}8sx~-jp{oeqAl@$XMVjj1X>wbtP!CrBEf33yXXVR3JZ!mrn zfp^p_hU%s{PxA?IRN#Wrmyc;E3QS1@xez^tmCUAkkI-5t3soag?9b}V?Cdo% zWyP(nd&Yp-hSKvdMP3+-2+T`$8~BvrHuf>$qz zd6g}5K49_LtzG!D5#2JzEX!{QdQ6V}sZvha{H)d`w%aV#;PKcCq?%au!Z+u$;Y|mP zD4BkMd|hv}yr&m@ihPX;*K+6v2J0K*F5Q1D|Gp3~Rb{Xs}m_~)F1^8aMW|WH5T&lgR z!sK_5B%=Hov)|%5oTEjUWto(M;pSL*lCBS{=OY+fN$VYMY;&peG`1DbyqYTSr<{+N z=ZE;@tQSCEu?tNuSFgXZ3#(t!eL<(*jDq$%g7X1ShX@|h`Vh0({4g6W!5Bk;DoE9c z&XvO~Id#!(aibp`LMIr%_3>(~LJSN~mx%rb6l?Ov>1ned#$r}SlYl0^Ukh}>2#hkr zAktP7*&mF0sQUicig(G57iwflii$7$MnHk~#ta2~%OAOVs1>+#r&=5|o&6??;|gA4 z(`ZD=!wWxYa_Ftz$2jk79^76>T;=W{PgBAZr2$wstC?C#TAt!%J0_w&90~J(S30;J zDgT?xBu1k+t+nW;x^hz`T?KO4h8vJ8z#KDsN(?xQ)Y5X}@mh2)J~k3RyX z-bjH71wN%}-QS+|Mgrc}hd7LMMiac))Sx!^L%2Dfh7`?+jS@DWtE-L4g&-pdtXOU0 z_%=!ADnw^X9q4AlT)mY_f&43ccBhL`a0mkqvKdq4jYUZcdQaF9`|B+xQIwq~^944x z_AW(_C$$FuNp`^WgbK?q-Aiz%lwH&W=g6cs;u2wS+3nsj<%K@ZP6v!RIXR`U6|Zgn zG}v0~1el5gbN|OFe|hw@cgzJ6?m62_O%4m>X=cknMx*!-U6Fg$X2~a8*Ca-eK=i`O zu!)W?mY2)caY-Qn~)%|a8TaQFis zOp57R@9hncaH#o{J)NihC?d=BlN)$)t)6LUYHrxtWmEdAea3$M!qPZN|AjsM9T=$d z+8b;R;ZGH9$@OjPG}?#qO4;Y#<0*EtSj+8soZ7z@{cn0I62{N0Vtmacg;DE|IOIphP1doB@$L4Z>px z_RQGSNfg3uG}0t0O?-Uu{WUc;dO>Y&_uH*-APV}wIRa4;k)ffg%zb|fTbh72=9fGI zd4t_|aAH&i**3jzCRGyl`9biVqoLAx#mcyF5gf_ElkEFO7%q(bW?5-m4#q=cD4aN+mXZ! zVD9b?r9Wgi%CC}Da`rp+?{q|fhbR^!({CcpqlC*rVD}Ef6 z6@6*vF3T2bLPuEG9B#Ky&KPPnZMQ*<&BfrdV&2yq6eLegt=P9)}e&GO~pQOr@>0g?lClr`Jg9Q}q^j)_ZlNj$1Sv2_mxGBie@h7*6VbY$E|3tk zNvkkX%`e=iR~^e)7ax|v?o2s1wYK~=R?9w~O1bjcM&phJkWS4IdWU$|d!-m}^e5=Y zG=+TD%oET1Nqa#YQ^TlUB{TvG*kS&aZ*DWfgPJw}+RMIQqU*-@r*eDaa`i)7f-J@@ z=CZYkL!STQaFC(85BB}DY}Xr&?EFcy*0jgEelz3%Pw7< zpn=rd(z3?%(9l@{ME7ko)b312TID)S)-aa0?K#5nBY-@dr|V%0Cl4H zNrb^GHbpt#v^b!_dPuXFX>L3@YDB0~xWNF~hej^_cj{}Vm!k>LG{yw!&VOE}gHxqR zG-(hzUY(O43=*=4_v?xRkC(w&u76Wl%RLO2L^$b2)lEHIth5<_>AHZMBv|JE783FF z_kl~4U5t6SL)rwQZwe-IozD_X5LQDb2?2psm!k|iIvF6`mb!Tp+ND%3<0iwUNl<47 z&kAaytE-#6y}hl06|w`oX!Ls_N5`<(Sbs}zO=|9sOq7&==jR7zdi|J5VzQU=UH~T* z3sq2Pj{Y|=h4l4Z&dntQwfSffbs`?FP&ETc$_~R^g@n_ z>}LP{_(91jGKwZjA{iPXIXp2^%K|ZG*}8{_yD(hOPc*)Hcm}FONSI}rR1!)M$j1d| zH?|wd)Bafm^4VvN^fRBe$o%J(g0FyV_y74P9KQd5zWM3DrXaige?I;H2!j8A^Ud|w zp;LsM_8N9~(f0&AJo8C$z@1qymVIZ zyK@8mSTW-N7W@3pEB{-}29+&?SH1JUJ8T}H*=u{Cw|SnWBLKmzQod3p5P1RZO7ABw zr;F2^ZhA*2C-QIt_rulmJ;)(oGt+2SsTEud=JDW_X;pt>C!6o&5uIWB{G zI0{Nzy|L1n-=qbh{b0}{AkYcU@e&5J^;Q}+w$Ge*YYAK`7JgrDF1$h#1m%6UPTDS@ z$mjN}Harjcn3Ih|KgSObD%zgd?z44JEgtgPJKFW+`ZE>vhupb7CCXH7250t?$PP1caVtx8XKv)3J7u+`On(JYtY@f*aI4G zP|@n!*m$fiVgS^0M&j0!#Y^Q^Uja_%q2Il$rsM@FT(ay!Uv!h+e0K0Aegxh95su z1QH{D9xv&Ex~$=RreTYLz^G?sRR~c7Y-{DpU`ULC@se7VKMxH2$2P%aBTID3c^?{iQfE-@-x6$i!g!`cKrG&0j+o5q^XX}l1#l=&uT3HB zbVVg$6O?f_YHj2j$7WJUYo6`766P%}r3-J6Nb2Qv|MzcuxaK)1Q(1+v zt(QEsa&|Kk-#S~XkNo}ua2tTspMfBPx;KP|Nu%{^=no^yG4N0#DT8B2a+*>1<^}+m z8Hg#f`OWDM;LY8k*%-jmWdwT#u%wv;(WlLHx8}-aS@xbNZYq|p zfFns=9b2EOuYBi;zhnjM(sToU6C0sAg%E!kd z9MOmO-_Bql5(%2@?p%mZH%lrd`N%}g!t$A&>@yqNqlzWms zuk2MOO*_!UY5J_l0=qae#t`J0(>qhj_aq>2Hgz^cLPGyeB$HNhjy8e zAODB8w+yN>>ib4j6r`m)M34sQ4hiY*me_Q6BP!k94bt7MAl=>FU7N16yzl2Z^PV~1 z&dk{#97Jc>>}y@uTK`{$=K6q4_W|!xmDXs-cqOsAuI}!+g$1S}fFxJ2F&8P*41B>8 z2iSpwo8(<#UkYVHN^!_JmOQYv#A6Sm%#J&OZtiZ7m=Eq${I{&9|1dwV8DBLUejw>hdP3=w z$!q_<6-|4ofSJK~&*9zA+jLUhnNYU8E?w>Zc-4IU40SHc&*$+PH%(c&yleBmFy>K@ zpZJf3nL__wE~8y0r4`zdcTlB7vc%u%yvgpgy*^rD+vv7e>AL?nDwGB|Be-1FlGkrQ z>l&q1O%Uo<&M}{>`_`S6<|`952dG#*q#CwVibfPa(C~r2Yy} zd?vBMAt4#8R-dy4NjchpkW9Jb^qTw7^?5BuoB3`*z@^*uE`eCa`IxKf>z?M`K+-?+ zg+=_x+JjENs9vM{Jup!pI{e7~h2P&4E(iHASi?!Fg)i~w+Aho>Bhc<*gOXb3!K81}=`4eNz95CWD>NR`a z?$bBb_omTEpO?&NKX0iq z-Nzkwr_>$IH9t43qZBF?Yl870g}p3?_q8yGZ_-74P?v=LMzJb`-1X}(ai9m<=XRTd zBU-i#_$(^i>S|V4UMl*YF9VI_<@r3&F5;%m4EMh@7k@TVvdq&> zGY687?7T5O&&F7&Jcn^d37ElX^}e>X_;45IadLvkY1}*0)zhH?kEwYpo-0DtX#Xc6 z?Pj|Ge%3``1_9@JYJURY(;6rcis-VJ>xQ|%jO4%ay0~Ll;Eswd4P$$QZ;C7ibU3Fu z0%5@)sADo^GuM_MDN_ZOfI(u**FVMOcqFxX9gvoiDa+ zv4CN|0FxoDvIhnKSAx(i%jMtROA9sTCEi2;4rmKc`joQ^<(25#U0=0O$n+Lq^2>KNaGZ)$F_?>7|Mny$Z% zmCl2jK}e)%W6$SKN&pCxeWFUdt(V8M%1&-iU&1w+^i`xF%&3+Ck!C_ zR1wU*a}5w1u(Bm)YwD(%Z2^XqEu-zhe?u$PzPQ zPv+HGuMui{Tfcz)4#!y$jOF2n3*u{s)@chrfBpp8K@E&o$UdV#&MU^0z$tAwmk{zZ zxn@7h8q^b|0JFk3o?I-fR3#T zB@q(=1;q+@m;3*3fAj7kx6dAaOd%D4Ua|iDq>d8wa3SkJuuHSho=RAJtO?H{o|#&# ztHW`!fKW>UpSv<3EIqnuzD}8&@9XaCxdi(Of~~xiy5p#(TAfPcgTH1M&#me(M`0i7 zn~O0NqO(0#pZ!C{N?umMSFaqY03a^jAg6q}C4<*6s=pB^eJihjpMm}CXx2+y^FyJ5 z-*{?JsrQT{;MpLao*lq%gF~-FnexqkNgMI~dKE(Jo2Is}= zFNTU;Q>yB!j0hl}fT}3DrTy_;n2Oha?J{U z{XRJ4sI|RT=1m``g0Z+s+yf+|^vA4#u5By#es;?!MbN8hF;VE_>iRR9bt8XJN@0mY z6!IECg77^x1}kJalXOxRz4wKnFi z@Zo4Q0%1_qQBhhh#N+0e;6B+(rl&S{Pd!!x0AUX40N^F@09WNN*!iz1{ljxg8Tre{ zv;z6E*mLhtD!m%uE&h0)#Ox52eDap+vk!hiXkv#EC5U)>CSVSiAmBbou086)D|(0E z*V!X&sOX1Hd#&l=pS3jZGCErmFiv$oRX6Zr9zInmf>@y5Yv%pOp?T*gk|n-P(1hym zvd{HXf(1cMVbD*)1d8#)U&d%W^i}3z%HT7S({kdf2;4ECb&-Cz53}VD%x&<@9x_n_ z`Lq{gK~H|X?9^B1VAh)~X_M|6FfrW`w^O4((%aIt*MEiW({Ar6I&-Cv4(rMG)<2` zP_|#LpF}b~abKOmw`d;4KLL^2C^(ot=?l_|rFV+`qZG zQMd{{k<;OB^t_BeM$2+FzI@IXuf#3^ zVZ+cY4&IW47wzD2A<54^W#=2UR;ZeXKB$#Sf@CNuSNF|_^QjbB+(unBqcQv8)aLPQ z4VcEWm%e`ii{j~ZI{L@gFBd?n zjzIhMw|d3nWMrPZ>pdYi)Q>rv=e7j=4fW6iAZFx>0hZ}mw)|>nacL=ijJ@NrN})D? zLcl>Ys-AKiE}QrH;)IyXCp^HjnFz(pgOQy_EuwO&XV%)3c!jUn6B}TSJt&qaNC@HK z;g`^xVuvj_4ms*yC!p@vYq!nnGmgGJOz@5M zV^_5_(K~c-xG7la-_US#8YsTUO{a5nV1?*TXf-DN#5e}(=J!54x^K@qLml8>Xuk6G z*DKR)AL42QCB+tMgO9Y4B{>th`f27WG_01(an;DD9@iNacq#wp6e0<~ntw=}^kS{j z0f6&?6K^i-wz?;QO7-Vsv#8H!@Km4uB<8QtWRI~zeizbWvoqh1^}G);A3S|7)oo8> z9Fdk*QvfLHo^~=VAO*0&lNCz)b)em^sW3k%p0O zf%WSSx=XrHJ)i#A!~XR~BSssn=zq!yka%3?)1 z*qJ^_KxMwm@E|^HcZHz-#N~u9m65$t|KeC<;>nO1q3g5JeFUe^9_&sM@Xvp(n^>95 z(wZACr7conU>rhI$l5JrNw`N@UsB}$1!&RX-eJk-B#)0qqJ_P4eU1gt=xg1y8_}~gzUf3X z7|wf|>Y7h04aP5b=l~CF!2A~+*2~fa#ODY9lfmbb(cRzb5^h8u!9v8u<*K3WlL4U)!OCI$FFV|eTR6M~MXiridN|zdx z=G*SteNh&N2E@#YAp<&lW^F1V{xu}IQ(0I?Jg%uPNBOyV9~l8YPf*DzK`BgY{^KL) zuSC?F;737m_N>OF*Zy{)uy7xM8uJ7+3j$u}R9(D+unQwAKOX(r5O1o;LUZx;uic#* zR^EGFl(#IjxGrx-Qv{#NEDcWVJywZMP=ETvffF}dqQRbWzZ3tgf+pC9_@*-eipSr; zVe&exN=`Xl4ss~{ajr4to{cbgME=(d^hCEOkIJ*!xXZHbOLct74XvQ<)4EgJ_zloU z%#>^WSW^HUsCY^ZFh=FN&t~Z3Ju?|BL*NqMPG{6uEd690xQj9IH4`iDSM$yL+?w18 zd%2zx#c%{jf^(UrDno&b^30h1!%8B|!Sz&4P$+H)=bGZhXS<0Blj#KoMbUFbl-_oZ6&scgYrnH%BoSP%Pqd5G zXA?gJu*5uV4--;eBN+T@@C>_;ivj+rN*VgoEN$y{T_V5)7jo`x6<>p79{noLc^Dw= z-bmTsA?dw^W!qY5axx#H`mfJ?pZ1DelJZj=`NG^>g&v5_>Z<<koak*lUN3EWEs}?uKrsk=J?fbN;BDDn;_i{HaH5 z03J6!Ebtu^(El6h-GA^Nhas_CbmERu)&UB14uq0wO^XA|FOp={)WlNp(#GN$wStd` z$x<@gHFUVwWhlO}W~IcH5f`xofK0U&v`h=a z+^%i(z~cWkIQtPG$GJl^Tb|N?ZL$0Jc(SdF|7o9Q@8<~{A@`bxqeV}_I1sA0)d_e% zK;vNXZ?X+x?wpj)of$&TQ;!x{`2bJ8=sG8P>G+y!#Z#egiGE{V-e{@qKGK2Lua7Mr z|I`0s0W`5nX(YIOEj~c{h1sJt7&0ulRHFC#Go|iHI+5~60`>+}lT6>JpE4_Cqc7(D z)90NOaPVGKy=v0f+gtaJQvtV_)LwL#^T~#FyA?o*p=MyD{^8CBIOODk2^q@jXCWNA zNIT?E>a)8ekKr?@qwMF`U$B4^hw^%u>A3*du41LPEcpKP_baC9p563c0hWncqvaE_ z{ENVoqi^9yTA$2!r%Hvp;qd>ZWN@|V0+afXgl#sd6!9 zU#+p;;{75V9UToOII<5pIoZ1y8wWVOMs#gutn=g?cLj<=`*F7)-@Jx>hQT{L)|oE~ zT~LaFN!Y=)Jkc4aAZIOiymWl|I75J! zrN8ER(bTy8bvz90lORwPw3MHnRjYyAH+552Tk(Qu0@?1H;fw6x<`m85?+h1!*s-ni z+T8%Drp~M>(ZQ=6f>cR$Y&m;)ocAoj*mg+9c{sX_Xfg#1-%yRF0#jwy`~4SGlXuE37Hf+EE2|nVfUCk2$mR>MWY1tg+2jZkN@5&z~%4 zAtz1&KYvLHyb|z!G7pRm?%~3aPaGTv()b^>wOACxtEsl1?L4glX$`9!M6Em%S>EC}qAX2QChgOAi7b$a>vF`_ z0_3U#YF5|?e$vk9#2~w0wR(`32N6X!fq~b{(|gPKgIm7>(3uSt(2EAKk%3?A0>?$o ziGa~|L*jPo0FLJ!T9z+gPB=J~0T2=gj8zBWFq*iKujG^xm@DLMu3_2Qyb4_>%XnkA zvhGR(9b3anL?6k>lEIGD?GI~U7cEF-0MdZbmoooKghstLXEg%h(hrcENPevuM%8gf z?u^ydQH=|=nj>CN#-hUGVcYLbSCBc}Ows1k?#EYMR-a-Yb_bnO}X!3dP0<`s*1p`P@7_BcmfE^eW~O zkD%CtmSw$h4?Zdi#tH(lM#s8m)Xsl#Di<6qQ6c2yPlw_Q>qei<=IVa3I0^p&s0%xD zM}hNwpLHKh3_V7u@w`L1w`P#rQdNa~v(B1T4@6SYt-+O+mZVUKMdxE%&|N9{-K|4@ zP4zt%-(+vPf-D_2Dju6(d4}wf{_t~A;2qE6eLmxkYq2ALCyReXxevaQrg|dt3(Y>x z+o2ti&sRl3@GmrkgtFzuc4n*)su*gfp12{v)hG&}5l*vkcm6ww=pMMD=F7 zokX-}!zyrTT;t4cULmivS$JSe<=r?kMAX)84=dPcSCg8cIO#>GZLS~wMNwC`CBf$w z09=~Cd8}Gb52cZ(pN|FYakm(Vd}i9eUF%FciSz#ydx)1MUr-(tA2wsX=(N`_Ljeyw za>_O0zJv_J*x1{uj#wIrimQE#gPz@skble~b1<3@1uqCA%}6XFjAB{;;A&qEF6TP-)Qxt} zOM@D9^~8OD+}5_|uj3ecW38>NR$b?LNvula>pA5a?dQZy>iY>>=8C#4b(HB-qYU=*Ia@a3@>eKv3CnM4U&?{P7 z&yE@vSX(PJMet%2`#ly%_y6IbR^6V;U(YL^-?jJB9U^G*rTM2wN(oMHz-4}O& zg-wDgH^_SrpJniS6)V(|#TZMvV}slV?-A=F7~~yB9TR5l|HGFL9yabDRrz^DM*)Y^ z!9^{1toMl7)Yx=-`RnZPme%PNG5}KnVwSfkj_NYg*2>M)XM&@rh5~n8qPC3IIb(yD zR>+`K)KI=c(RBtUQ=5sZ%Ko`_?~Cn)*~La)T$P$NN#*0g#(^CFOQkst8(Ui{y`SP~ z!SBJ9RY~fmJy64dIWjUTu+g_B#Kto*OoI1vf_c%KwGb882)~q!+k0n=sT4}g9fHju z?`ZC8EVIuBMmrYdxDGSnOyBZ`b1?32;nBD^sO9T*et#HJx14S9*q&qxRfN(4PpcxK zoQ7oBYy6w#!<%y%rwJIKlZK2d0X;Hjk6x=?!ch6IsGuOb!f4S9KA%!fv%X+t9qu~i z4?n+s8c0&9g^L9~FfJn)d5F=8b~q%u!4|KN9>6A=8eYi znIFQo$z+b_$ypqQ;#|(TF$%Sd-DzO3PVASiVEdjd2JNOY*&n3fYHsivLUu~!>&SOn zVvD()&%d8dle7=Puc$IOsHx57ieyq_s8&)KZ5Z}ulBXIELghnO z_b(Sh6@lnSHI(^N-w;Cu%!z<)yF=`RsOXyvAg*J=E7RPv_l%*M{!O(7QUI-XJ`YQ< z6{*IC&6sw^B}UGB>u%F)f4~ljBYz8$ELuEaRwf^qDRqYlWj*^q9$fEf&hM2$H}jg6kiU_IH!Ht9PZK)_6C-G&$;*ZAhrht{eIfVl;+4Hjw9X>Rqt z*MBLOG{NZQ{)wLU`RBZ_xzX!95=`$izyD{4Sg4;M6o0MATD5- zh1~=N$1EAxFiwZ~A=}w%Hm-VMS*D`DlubxMLP;m=^;#o>HB{NoWQxZLS|wg3h!R`t8* z&FT%0ZhoDXUA=SRygZk?rl<0P`CJloL~jtD9OPSjWZUoNCIiB2r4*3!)_;9iZnmwI z+v^T*1v`P$A(eWMa{YF}UAxj9(Zp3MsOF*T7U(^%+YEuXeYwi(S&hv+?RPuq$y=FQ zX8y>)#LBGt&06CnY0fAzB`X;_JO5bMGJ=itgW)^V$O7$3q#ak936Fj~{kU`3e_?&8 zu`K!39k@8du^=H|O;w-P3uH}4w22HOMm6KlHyM&E5dIwA1btLwE%sF>Hcc*boI z#Q!lq7!3&f=J5LxNYgJ1Zmh)yl&6|Ut-kV%^{O9%< zzStvS+TQSdYoiSEC1iiuCD&l5#POMs-?*rMpEp?X5hnn}q7UJ6Pi)3=mdSZM70_p+ zX658YNYCgxIz(bxh*B&Yy@{>}9P*=|sU*vHFkbX-?_LRUBvtgz_9niTaxjF`LaceZ zIr+ROo((c4J!3STcAKrIT40?lb{}=*zubla-p)RrSwX+h00tT4oRIi{$E=JSCC#Gh z)3E10u=l0dL<3Vg>`)W*aJtN~Z@rOsDLW?z)IK}z?GM34C~Et7Ku@1MD}(;|@XL{Q zXG)cW$Uw@kj5Ma_<+-5M}cq>u+5_W-HNYUxR-Cjq(8vB4OE)s~heUyzQPJ*SL^3qh2D=iQ-G zN_?0kCgb7W+WzI^&IOJQ3W1`+ zAFF?|G_Ai|k37Ms_LmI)B!kwkpK)=bf`X?xlq0}gU$psqEKceVe`Ay=!p;r=69Kaf z7?_-92*{*!*t9*rM)kbBdaT^oQJR?(**{Ji(Q6~(HHcmmM}Tc?YLW(1hwtH^+1_vb zXF=a@UhAtXr~0Ossi%tIy}0Up@gf&{NqQD_N%}QeB`9+DYpkq%zZ`dCviI{#7ZV-> ztI&B)`mpAtz&?Xo~r;NV~wet7;3%&5y$W_6_?TnJbZV+kbul9YLS zIS{uL!Qs?EUtcn>*1;sSzV>knhJ??(?u1*MiI4BD*wQ>o!dkaXbBTT5QMnK>ZXAkz zjfc3JJKSO1@(Q0^x&CbZFBTAEJXNI3+2Yb_uyrwc#Akm#*rod@#*!?YG=i9a)f-;8 zkc(l0vVU=~z|KxHdc?~o`sG(iXp4-D+H5>BW25(PcUU7)8R`2AU3&=LoG_^_G0^zmGzTtMSTj}k-(m7 zTQMiS>c2*V9}hYMJu~gDjqage9&ezdlge%F-zJw>z&_FlsH`ri+c;bhfK-@#GC|+P%+R(V)#2qRz@cR=*u@X;A+Bw_L);fvSFLx%f zE&!NWPwHom|MSUQNc}!*YvE!YvJ`{|BQGQzA+JX+sV54H^ zo*#8U0CCj0fyT2ICsw z&hZ?>o`+v)@ur$a`hNJ08HUR0^^0%Hf&|Fz@2+x`TOnY@Ynn|Us!iFB)CP8XsW|5h>m33x-ed{6%E_9$x80y27c zC!2PsM}v=6;Y72832>$Wn@R|$Pl_dYXeX{MQZrzd*nq8#IF>u}*s}*w51li#2r-Mq zJ8eY8Dp%SCi>Y`Um|CvDe$Iw<2znxB%z_!Xk+ap;z;Hb7`Xih7HvOph=DZ%ynPPk` zve-TjJ7xOUcEl^uT~)gDKWa}SGyH+}ARB`5_exqtHdssmKo`$q|ED_7eG7yK%0=>b z`?o(lk9G<-c70_4jJlS>?J5|7$wzJens3EMqn%Tn{P0EQx~h9z&Ye)R(ET2$RI_=*f`ZQK!b0|JSqg>l8i7!? zqj6S8OEq^_R_Jq6DM9>bsPr}-S4sIgkDht<=%$~N2aso#lgo5#2gSDPfvDM>_gnq3 zi!IIy!Nyq|-p4l#ZSJgpyXK(pv?Hjm40q!PYATy~?N7%5Rc;p=w|1p?exc;>(I7e72krd#{#;)8aM zTv9d3gY^FihLAMSqT}P^qtOGF*DrjXaM#Oqc0tiqLr6+l+JzUC*SW0k1Qnt|)f0zB z$#lvoEG%qs!qBht2TaI=X`C*9lL(hXo1#U(^m3bh?6gRjA5%?{8Y3YTYHP@>wW6Db za;T}5r`U)&#OC;QBqbpnlP;4?E=6C>;8^?kU$li&prkCwJpqE|lI`(bc8|-&mvQB| zqii0xhu%$d%;uRBKMWcQ1Ha|3r_X`-+hKYl4&#b4hIDN3nUTIZ-6b6vn(|{NmOD;6 zSjn&MH{p>Kl7%Fy*cX;+}=Run# zne&vdMB&SHID7d)6gW#*SR7Upu+EjL^InV{vB2Cr#ng3<1QDB1lYeex3mz*-%9q(& zjuG`m>i1OSm+aFVyVK@!`8LG2PhiWT2X>(M9Nnoq>ZkO+sG@kgGld085b;1~BFGU{ z1)IC8528cu@N0=qfcpb+*GzMDG#$oP24mVXj}r*pkHBI{kNri>=*qEfrZ~H}sJ-VY zo%P*~dgA76YiMqwhh~I%OyjdyRYaT;68OIuEO{j~=Np?@Y)S8j;CTov;qSYnxO; z_JV}-MYgRNv-U9wHm`HjxTAE*yp=erkm1c8R0g108T$z+Re2E`kTWFdES{!0Mde@R zv49Xq2}=vq7&vk#M8L5ZtrRZHuDLyp95T1E)jYf9E)pF#?(vamqVt};pW@fjY`cre zKTka=#puCL1i%b{9vmV}Onhu&Gz<*kTQU|2ig7~gvVZv#6>~q7fnVeK>M@|ayd9vU zs3dxZ9}G1cfKjlz+7PHDm0H!EM0)-%w_fk=UCe-QW3pu;+Yjv>t{tnkn0^l9paHU& z8OD{CS3iVcUVY#FCz_oHuRqImATR>x`X=P8JgUtWJ!({eCE@h!GRHMcocvd)Z%VNP zu&_>z(B*OX3Z6b}E!gM^eE%*Cj0zysa~)&d%F2=oSo6^Pvvaq^;r%)Q`~VJ^M5EX-|=WHgIsJ>YZs zKEiz0ZFd7cqF9rXy?RD#FV-ezfPK~b>74!}dpJs7NK8(3wQ$}!1jM8yPG&2G zUv=I`WXObuZD63=H*rqzarr#);+H(dXL}ecDvqasK11*ObaGi5h3r5|-Q-^7q~(Lx zFMbQaw#pje{xjZ?YDq~^*1p(7_~-N-`j_^Q=^+Kyz`xh%fBwgRz7zx_;D5dVe}w(F zjqBfE|36d=_D-{x>Au4xB;yz#{r$VU8>pm#jVbm&=C-O@K(*|FQz-b3|Nq1ju6^u zOFl>Fq*^B&4y~fRXR!L*@Nk`6M7ROb8MYgbtZeu*!6g=UY@z6Sr3KJ*&&)A)dXSd` zy~CEvU3=2@T2)0PQm)319I&Kru13m3@SB{?r>e}nf&zEhy0HKA!MODnR^vFI2$-8# zEL2*txNszR!A0nuHS0sn<{CZ|J*aD#nP>fFmONRfcdZm&Fa~fB5&}vRtD$~!ij1u4 zzyOGhS!_Ldi@iTxw!L=jybis^vZ!C50n#1AY5;O{GWTDpe?byzx$lzMt0M%Iy*D`Q zY;0g)Gw5PqK1h2@6{3rZDyFcz3+v*oy(j@w+tumVL}4}dQuDL;xZ&4`NaM;OP=@f! zrGZ(p+WL-Qy4VkyOhUa(HIFwQV7BCFXsvCt*}mY)QlYv*IbS@po2fTzY&6wcr~(8m zwX`%WE)G)8v9NX-l2GpFCl&}v`q8S>tQuDN?i16^pH9m6BOSbpI>?@Xfg<_rWEZV- zX;J$f(T)CO2BmD;$waFmP|uz@Ts>et6S$bG7Eczr!Sr#KCbQYv+gDUqcg2_AuUwrw z;*@ig=t)_#kviBr?p_*u&{-}x*#d9^d_MTh1@*(+uvY~jcq}{2-?x(j5^F>H^i+LjRfdaktBjtFW6OLi~nUZm6?@wl(k zVfQCc7n0N!s5RriLPjOwi(;-8tKMO(uB+@88u)by0z@GlYv2a>pEnX|Ad299-v#G! zMbdigbN_em;^y|W)jN(Rmi;uoU^j8aJA*pjqkT>{uj*5Uy@4Y7Q*B*U?4Zdw)XyO6 zOKtEQ&(^IOG)D12xkv%!fILW(*{2FI_`*((f~jQUcOMBE$jA&b|9jOMy*GxdM=N~2 znv0Z+CUzL;wOh6)lMwx-hC!;3Z~#&wn~jF6h*2N9Xje$K&!5MDh=Bz9ppM(>QCrJ* zlE1BR$qHD9N0wVy zr+Wbr>HI!@<;wf{hyZJzJmG%N;8aX9kYa zhamW(>}Kg*r(m7U(%)V_bmwC}y7f2eSs)~NkEjf3f-avvBSEY*mWeSa@EqZPFEY>% z^1GvhXqwuCr+S6Th)YblUON@nFM{{+DW_ zS3YboWSyo7nuv*jnqlNAC)!v!5W(hh&*1C8@NjsBC#UUGwhCNCYDYOJKmD*CN@s_}&kT)W5?@Yx{Z1JQo5Y0N{@k~jxTe``}$-cO633afNXwtP8|L1vccGiyb z(`NwdI$rnt=H_N^f$7d9JY6*zFWLB?Kg*IffIxin}!Q=q-Xbrd8u+jt^ckCt^yAZso0m=scGY$L|Jf5 zXFtyf507#Q`+l$WSDJk!>;LlKMFZAN)ZC|0!T;HWj5`Hkz_1aw#nJ0y=uj67{!Uk7tA!;EjwqRGz_CR>Cvc!Gi#X{Y|4d zKfhm=U$deE1Eu`}A*tMGpQRS%(9vVSy>n<;a%fpf;V#4U-(Ly8(heT++4=?!GBQaJ;bh>VVG&^CZR{J3*mhb3`Tl2fZDEioML`Mc_+A{71|)YPBG7Tt zKr@KVhN(?VWFjUpRtO0rhge*2)X)rqoQX4t>AdpIF##Hiyv0Ka!O>$F%J0G*xg~$> z-Xe#HOG-wcr_X$#+|=l;=^4@H1<|x3x$i*c@bzdPP#=qzf9Ui`NKFmBcXO035or{$ z^z4Yqw59YQK*fOrr6ap>e~{toi55MH;rx2KC?%oz`XT~M3_BmU$*=CoqZ6vKDUjL#=<7%GB{MGJLI;td>F})_&8U2 zz3XpvzrN1aWeMmya9j}Tm#X5Rl%IZVw!g^QFsVE}Yi_krEhXKe2vM_IcDcdCZX>>O zZXBl|SH!)%FNrisetVda({AtJ6tBy;cSO|2hr_V2>T{5O!-lGh*{C3GDJ0@p647%I zFBL>pUr|}tK~F_aH$T@moW4Y*i-W|PtG~l9)2K!lajTzLF2_%ecshxNh8O0yE{dQ@ zZ4M92ret9IZGGQrzi(8lF4!6q`$q27t1cn=Qx;`gE-r5IB*&YZ#yA3JaW{2$uC=Sy zfS@1=&??Y{fD=gt*L=TPsovmt`9+sjsc>q$F~&$0IPcT$BPv$+*Ef&0-a|5=is&R6 z-)OulJvXjeHK%AK8cw6=-Wmb7~kq6FSd^l<{%j+N9i+^3BW^e}d z)avTYi#MvyDR$Au(>qN)s}}!qk_lw~oi0rI_m!lnRr1vJxqZTUBGkSkQD8?r85^Qn z{q!J_O<1BfGp)2k=ipdlGgXr=QvSNb0Vruv5###fo*r^uDqcx2QUFV&&ODp&V3YD$ z#}sWsjD0DqjR?3X-%-f=zg^y6@5|kJt{|=wHBsZpwd_bcbSO zbilnQfP)o?86@Zacm>wro>!UgYXW|{S$uX{Zf{z^GOcR96%UiZoBir4f?4(|Z6X*e zvi{6nNT}%t(po5pYsQ3A!ZTq#{mhvQJmkM0-w(CZ=;v2J7Riw6DLD*P6JmaP{c-Ct zlRT-*bkH%3U>=m>yi}KYCd6axl%ne_miC$ZeVOg>>n9+T2#i&uoe>hK6}_5{iY(Z3@D25qPT5Ad~9thpQt=Qj({*w>{S@b;w($)%&|}&SL5jIaxMg_fQu<6;uVf9`6} z1Lg{5%)Et2kJ}u)GV}AZo>yz%faq_dtBH-aU7#Z1so(Yc?fnW{KT|!Z2@XIyRA>@G zJ3IJUhbb2mfAg|fneP5c%O-n~Ax}!OcTLMvUP-HGu^iXn^?SD{+d?xL94W(}UA-W}|` zCRBW6R=@1u3P3*ofal`+hH#S)POfaFNvu&5)38bbiy^=L;;NN() z-aFY)qP`AP{M&oR0>wbecsdm@V%e;0f5xZQRWRQ8Cz<27lRa!m$iL5QORJP2O!ZN- zju%F9eVn+y@6(B_(Cch_lo^jTKX0F?AAq_Ih`z)nQOI}i`eWnNGrdi(;d1q$V&}Aa zIX$$0xjQL2_XR#HEbhRCY9mP9N`9rRVRut4mSnQo;=oO6I>bP6g=3-K^1QnrnaBB% zFIZ2&t=*gS{(*t`0Vog8Fot(^)>bvuFH%-#?7;%8)UC^i^aUveyT6|<*t zI!HV@j-d%#lT|`yvT9RY4t%iNe*FD}XWag)D(aa}M!H?w&ZckCU<17)XV&a{pnXB) zsVR;aPjkX2^&cxoFCMg-0uy367TvPl*spAd{Eas($4*~iX8&cGBVeOVpw_^Q_^8hB zc3f59q}F76J0Do7M6CWy;ks%sR9KYUf{s4e=Oc88pUbC$xNsF=mVx z^O=-LSwo7+#dR9#?F=btusl*tn|k-Px8=DS5YXT6$izv(S@Jk_JwK%IKuw!m>4mhXIAE8?=OBGUw1u8jJEH%}@CTdB zqn1)~G}QkY^I0w5g>^D{Cah(Yl!_dhqJV++q-rb=4`q2cfS7|>XQ&*F`(IeMPf6R+m;_-kzsuVe#!WfWtwzqg}m*wk)`hWZoyab zZnVTy3ZEydr7U>eR(x(S*+g9jh3qWOiiSwq7qoWhshTohELO_IBqlaIipd5OBNy;= z{794SGE0_CZ~fU%))qQD^Z7?TTc(YnroltHT zyrSOdQeP*EcuRIBHDfKx2~)TPB4JX9G_$Lq7pQQnbB_Bf1GeM}wSMzoTe_w+SX!}B zez|TsIo6O&)=Pm$`oBZY$axsm&Q~Q6u28C!6K#a9AYP`^YH&v8CvUU8ziw&t{cOfO zZXjFx=bL9O7tt;@l)#6Eh!U}h~118r1w6-Z!m9@ z*-C!+by5xo_dlo*xRCKuCLwUufjC-vf|;(MWuHw}Nip$Fu)Uq-hg;#}44(X;^CeBl zrMD*(aLBT2P42J%(6cV6-?Oxs8j;FLsd*4+Kww_CP-}{d`(XJTrIbC}=rr@^ik={t zgF&NOlunsWt4QKUavl5h4x=89jY#F|nUX3zw7^Kh=wKQ>#6BVGwpC6&gvQ`6J5aA>txpK>UhQ3-pd zZFg+7Hrz2<#@<6M*Q(O=melywqmv7t@1yFzZ8jtqcpkbMZBlZPv{)k&Hc z<$tk&RhR7tNZMBm_t3!_8)r+m^F@l(j}J~Ge`4rUT6~qiN~z`>s^^k>{9 z^E)TuyUj3_@$XdE@Aw?PWmrAFE56t=5)BjuXy6 z2=A+_ld&vQeO~K3YbQSPSIAofm}&72sLZg-%MC zUuohP!vr*nlwHxDX;P2}u=I#D0pHNrl1pFd!*@s&YbI)=O^w)gYxFL>2p+ynvoiH9 zD``0CD`Xt3KW{cO#N=LydKjab{&u5Fnt#ypA2JX-rKI-i|G3L7)_pTcW=U>SD)Fcp z{p;uId0%83*oeL;^3|B|Dn;&Woo#|69^?E}&f(7C?hUy-#)(axz#=~Y8pn`elM~(Ry5GqjGWA?w-duCQR7G_!xxsLi;`Nob& zUkpSg+66{N$-nr&D0{1@D!6cM^ix75r6r_0rBjebIs^m+q**jbcXxM4cT0EoLb|)V zyWyYiJ@&ajT*DZPHDekn{aXt|BwYPy3R^RJmm|Nqqbhxos zPw14SBEr=zUGkUaE0tsN<})zl=a8R&`hW6X6Gb&uhu&Z7S;wOVi2B|64BSea z8W?b2w(#pF`(Pc^fp{G2LFmPati+n!at0($H&5XW{5>y3`n$p>N`NkS7CGl}GPCrVS-sC{&L^Hsu4 zGSjZXo6XMTn+U&Pfdu2SK6V*gvt~8>>>qROV#+Id9coCWyfDq6{Bg86Y=KI{Z?56* z79}6*TV<~#Nk`v*#K=zPKr?ucH&sf5&`xa{MD3H)5*|4vn(e=xfdH|4$$+g_5=|>a!#k6TN|0vQd$nx>JqHR50l%@eP%bx9Ii7+4VO2bS5_3FM_i9q7Hd{&e7wRwYKk;;@>)g zr#L?{()gW{dpgei+`5-EEMBN9$}~}1sq5Z4>F6mr@pr0G`GPOK8wvxc2Z6(Z2>MBj z+C0Rcx~(!^)o4hmm?qDY{y2ZAhNBjXXH+Y(xl?ppswNIEu@bb5!3gE~{QGU}f&x<4 zT2I$8`_XqK$Bc*&xn}V40C$AESYwiUe1SSm7{ZsXQgs%6SCb1+8l#BeReMH_y>IVff8C0r9H3=k!%3_e%I{40_nSSnlifF+D>{G3yY zpRg9LTI7hPpu$F6b>n-Ss>mZB$hN_Cc{r{P&AA-LWDG!5Ez`beyV(6IZ)=7$>~3pY zqUEi0U1v6+%i}?-P$K`X#ZXeL(cMM!s?Mub0vY_x>=BNXz@dWgxll{!XP_QHuox)!#dscT-UUt(3Z&O)MiRbj%85 zmwclzwos$5*p}oorvZcV7GZ;K<-3mo;ptH@$mbhTYe^qj=%~5A(eDZ%HRQHtX-ii+ zTj$soOWQYm!`@WX{z^0K=YQZHpNn>-d9D>$*faj)jrhW;m=!EQLv8054RjG5Z^f7i z1sq)u=Njx5u#w4?)Qae^#_}!yE-(LVScsuvUoQM`(p`(BRenrCaYCrIh#DT@*beG0m zzHR?tgZgc=C8b0JpEbmJM}l!N`Ezk!Z;xD3m54_x zK8&pe-u(mh2nzN5AL@ptgXJ%m2G?nxuo*z==P?tgt-)Ss)QJH$FT$MN_+byOYzejM zZA#q_JbKG%?yw;qUfy0f(|I>vaS<9l6CMXgvR@hR=GxB-OC@JRA$o1Wz7o>7es&7uNuzY=i;!}VXN)0Y7i6ol4 zI&czF(q?v1V`O6!nqySGko*o+s&6W$VI^+NfEri(+X3eFH?BT-=e#t)sPc2h1afHZsBOoI7Q zv=70Es_(Mb+0~#>v=dy8U`4Ki%t-LeY~A+I;M?8E6rkjI2Co*0q#VV<#%j~~Y8^0m z{prvspvS_>DR}j~R|jEzxm8eGok`adDkB*ovHBWvMoD4b7rd_w&9ne+7Q!CZF`gy; zVOf-Q+Gk_=ayB1gC`!Qg(0+)D8D@{v?-I!^sRSk^0s{_lmxa9xl@CO6Uo9Q18trtE zN76*~Z?FF@V??99ZIm!NRkT`sMo#mWs(-Vx@jm*c)3^`tB zhq7TLi$>wgzfU2477dL>e$)kZDAGtiJNfG|Fm1{pe* z&)?BJM#scxYW}_o$R4pmqD^zpT* zC3G_t4?*xL@%f8$k~ccuc-~j7H$&C02d6DCz?<`T!_Qu-9%+gX`C7I%y$a}LaDj%1 z)cAq!2>lG4)iQeHhn#1@N_3e0>8Jt)eG`9UoHzeg)+Dd(0uNx&d0gCA;qYw!T7bUm z%03LbbLKs_AZheQyp+0t(KRlmWwH5}++Er2tuGSqMVszJs7R!8Vb>6Lri4$VT#TkxT#hsQZy-P)m&CQ_>kua-AwPAw2!XJe5gM;6)=!fw^Dk zUQaH^X;XV>Mw~G>bq_FjS@d6}rMt7gB`NmE5W=U#S}pXk*2>C#|BAs_*B@`m$bxu1 zI?)bO;8<2uQ?WNyBhwU*94GLZH{TYOKa%Gzm(OTzKddRXMmWX(FrC?>=Pw-H%}00s zkBvXloenJEFWQx9(^;zKpULv$uDAWb{}`_&9&}G1ZI((c_ch;7EU7K z^fMV~fixRY-W>pO*%=y$n~xJsJdgb)AwoBCBm;{)nc0JawqvtMfY0Mb;xiiVb+cpP zyRstlX3w9I-_?eZrszz!GiO~+b9Utv7)sqSl1QQW{Xd~{ocuKfwB5|z5sR*PmS%IC z$O>d(B8IiI5pUco1er~O z>jk?noJQE^S7;wYyuH(doRAgk-hT$zSAEw86(fI#Z*2d@33ZB@MDoTTVfy|69=g9N z$m7ju2Ek`I>RSjj!i|JH?${kTcXo^2?&;U4C=UXIoC8eIl~PX2w$r71+BSZhLVX#E z%sZkN*%ya~lLu|i6Bc886BukL_FDr`?jrA0o|m8ll6-~vUkvW%85EBL=2aRxr`R&r z&<1BL*lh@$-4e>QJdAcx!Hib)LIsP56mW)i*5J5xfBiY&<|EHQq=;-6HZop-VtE;W zop$SfF)EZqCc3=dqMmeHo7oQAFIlL3gW4-ZwEcfrfEKlQ7=b|cuLVOeap9ET4o1CY zPRUr^TBF1hDOs*Yyn^^4-UlY3U18Dv$S^C^b0Jpe|Y^IZW$Jtj#kI)GoY z?j$A7e0;;&)qs_)r#t+6hHoYN=X9~uK_{Hs$)ZQS-Hp!KOmHv>?Tc3ReXD(CR|Ndk z_0?H43scp*-aTGtmMy1`5v%7vZv*Kw%$O3IF2(2m(l@l+PqrnHij%kypWvpe;e5Wu zMQA)DHUWC@u8zOkE$d{Kb!ync_1e17m@}Yi%{AD*bYMj#n>R%SqW?VF$LdK2mk71M z;4iViiFH}ZRgJ^p;g0t!6(m%vCSZ!h{`F~l^Q3BEd?>SzJ-jA zvul{DmdD1%*6W>AfC=uP1Y%$?2M|N{6nh=qt*ItdJQZrC^^WZSFhhIxx_YQrFj(TE z(=^qC26%5wsJ^&m8ydvkT@!u&MPrw>n3lRAp=P*j7P%cpt)%hkSqRYl9)LIof!ASw zIdl(W!5{A2TsPZa3?Dzye5Q|$VRXu^jpcSYqm?xVD52>$Mk75v^G9~kZFaX~mNzaB zL3*t;CL$*ibH6=QhO{fMZ-))YdS%^jclSE?+P;=Y0oKR;e#tV#Fnq!lM>?s^^k@+& z{}>wSMFzADE$Qwl&OqzD_0#fZP8{aUleR_r7TfXk^lJjh{ho<3#4Lj8b9OK_O1PiO z{+4ZW2Y9Qh=RXM;W{b6|20Br3XNNxCX)?)IOeWGT();V0&Bkjsxgl)|**26bR3!DM zSN#Iy<`E6MiL>x_#Ua7IYKSU-pB#qvTIkyFpKvO7*XK4UpeRUe#_F0`=OXlUJ1O$f zJ856qNz*} zZQR%8+>S-UPRgMv48jAWzL_(u->OVR`C@`tJpWGh6iV?dARuD4 z{Eb9$F~aXk*Hzm=9Uu6POo2SHgu$?1KG3a7PwH+47&?Q+UrIu?q6%TNQM9#GWPhrO zF@O2o>QcEJTc%PIoCU+$x*v(VsnREwhfU&B+Js7i?Z1&d0rzp)qqwO#6n&$sFS%oR z2^rq=bZhL7#Z^}y``K4<`UG2X9(z8_CD40#(IP?-Duc^u2?8Y?~_viP!x-t$LE zB4=B|!qJ_43s+0VfABHp=0P^OMTD1Y_hTya9l)QKss$zK#}Cz7%rD?|KZtI=e+9<+}}HIvAeQK;zZ4E@Q;^T*s6%2eN+sea$a$q__AFm#Cu1r;x!z&rpnnnr29vM^9 z;vp6s3^X1`Q#I_bmxDJ8i&SY|79zUh1aB_+={`%GCJ3X6TV`Q2h!5yY))5WL{1&C?7%unJCZ4d=uYjP1n4626q6uaZzKrER`q?JmYpdY zmtIb?e@;N&U%>LCPBmnw7RtDt$y6r0!}?A60}iSw#<1&<-eRO@n!4PIA5(LQ`pr4u z_|8zh(ocHz;Z1zKTiM3<^*&aO2Px9d+Jyk)PMPNr9X zzJ1jwYXTPiS~h+7oX6+h9u2Yn>3+qTpS*jAs8zYbjK+zn+k1`m%A(D}yn2Xv^IPU? zqua>md&>^Neobn&Z6bU^FLDg%jtvvZ)3Ffzvwlu@cQ>d{93UVi&9>!;MR1#yJ&b`e?lSiv=kx3C%-_K(UGwhqh!A8R+APDwZ-CN6xsRe_7L<$=YVoOu4`BF)mK8l*$pE(lzf|x1D zuoy(&Bua`%0J)otSY;xK<~^0$sZkryIrZe8PBP+BeJyo=B>zP7^(bb9Y)GPQmW2f? z%?07^K9OU8lg=k6i}WM!`o7uo2`%5qXIBDG4Om9TuB9lT2e950?ji*Ot)jolVPvG) zU7ZzbOfe&cb4(SEM^^7RIFhK>e0}{l2u(MeKRJd6g$j8ey*Fa!YtqkC^p4KPn2GKS4|#x4t5%oH!!VNUqE}Rn+V13*Qfns9ybewB)yS<>nRwY= zQpUGGoDR<#uXit9BWkA5HF7u>xHsG++7NeFXf~E)N)zmIKU^+U?&FYAv0SvCJE08L z0b#v&CbLcdGz@hHoRUl0$SymT=`Vm*UtqI5y~&z#%#|+8iZNs?`jPp!;KV;zt;KS& zwNB??8AhB%^S|n=KaX*?~x zq6vc?`Q63>QXgNQ9{qN8Y}fz#t*u0K$!eZ8G3)FsCFma6bd5{uCpv$@m)*+Rm1lN7i$ zY6i^~I1OP~r=KT`T+!!Z;lH`+eyv0a_YG$#>d1o)s=z0hw9FVdB%4+vK@bMOY&Wig zIywHK$ZD>hU(ae-SXf~i*yBtVV=M0WWyXJ4&pjr}^1@S|yLL_vs@H*fV>b&6A9P9u z9Nv#LxdH76Ex!f8a7{T++&z9qfQf{Zla%h-smcvJQ(2^6>c4uVH)Y_iILtl(gr&94 z;=&?lf8*wUwLmht)=)0*uNJ-7_b*Wo7ZJ5ma;9zlg7-J|=oUlwS zwPg_YkS9vHV1+SB1*?UhZzorkT;(U&3{~?AELS9%(hOhYeWtybIG#28yqpw=R z6>Lgsja}nRVN{_gBZC(Ck!%&oh+^Yc)pA8y7Cb^3+amp|wUntt;oxR+i6X~ZM_j8Z ze5a64;Hjxv?zqazqD|1wUm~Iv1AxQ+Wn@Z|Gcmdsx3j zQsZdo3wpQV#DfL9l}r?F4;49ZF}y|W3D(!Oh3|SVaDSfv?#}J}iGRO1akjn!!{~+C ztz#Cg>nz{*78djO<}0NbpbEuV1eKp(-RTw|d^bX6?OkFVBjtmDg?1`n6L>F;>ocXExk zKy=m<3DW5GwmjlX#XGhE5rTn^v9$CfN(#6EDoi$pfqy_9GMEKY@MR?>X|*N`HasV@ z>M0;p^#jhTFAouEDvlMC_cu9XZ&4(Z2lAU02$zAf2($pZr}M1L%|UM8z@+(!1&&|m z9X)uYH?9tamo68mIRUa<@A>!E+Oc|@Ec;hp5u)MO2#?fXvf8;#I+0)vM3^=Juaz0u zY(4`jJD&n_>`h`$yx?mgo?cbNviGQZly$~GPWqSx7OVAMP}^kZTP(Evvx9JF;anH< zSHgQb^NEnGtZjsBR~NLOF}+e)C_;4dA0I6@LT?V1JCu33turS{#$D{pLx>gh^)CY} zVzTwtV~GM@zl?0;aTF-IeHoF(Df{tazD^Kp)59ntDz07cirhyO6%N#Kk4?H{oyXK> zQn-6_4KhvG$O4Zxmmp6iJe_}~RF4jnCM))^A1Nxtcu8Qd6R#i>{Z5Rn2XkDx$Ftt^ zq_(KB-sJ5b$SV@webr`iW06G=AXKvX*CPQ98$_d?-ZQv>>{l9i<9^-v-sLrUGSWXf z2d$|5M1JityB5EfvB)9wE6P}P_Z1$)D9A24f3zwbu(6An{81_^tAT>}35Trrg>{3o z!hxuHiB3(D@MdNOAg8A3^99#iqDKR4NNnvki5dvi+@mp`ljvp?Rq`P= z9RG&}z?t?RCQb!W8wstM@6NPkonK!>z)uf6ezz!?!J4?tcqGbCq<+1%XWD@jeX|Fz>B&B zP>%8I4KYpAq-asm5FcKErK2mOuw{^wy*7WYJ0_)23GS2*Pzjs1nD<-%F;s*F%wi5i zT-UN~Wj21?&|;6!XFO{sP*O+F)yzd@A*g*)A>e*Z&xY9}F<`v+k@;KOftfZ6N)}tV z`VHNTMu~*0Sb>Xb1xALq?D9BgM-t1WM2tcY0f$&z$46j zu|D89>72O)hQlyEPtVt{wNh$N9*pHJxusms-CxA!Nk7vr_sQJzexioO7jJRP&;7J}k!x-ZZ8v`<59%@_&5J$0srIsdyIW^27+Dc) zuTt;j*4S+gAAbmZ8C$xjL=c1lnN7(TST ziO7ze&T&6_OfswdRw5xdh9ZU4hLM7dyTB+B}7`+PMOX zxn=IlaZSJ#t~dvpya3ju!D&wXf~8%8d@vnZEeb~TiQSiI9d0#kXK^IBhVJgrHzK1a zt^q$^Ayi}J4JT+fr0_`61ut^(52{!S=1|g7!cX1m%2y%b&gQ+sP4mk`HVuKZhJEeh z&AOQ9rRi7_wmMUgJbv~nAKjXW#`KhiA$c+;WGj${c>DYolHq#2rLZ%YFvTYD4&j|> z_|hLcqE4NA4Sa!fL|@3J_L){{WM4K_azlarF>mG0s`b|iF)@q>N4yX8!RyMjvD{AA z#C1huMX527mKcAN4-jRB;lMj&Zw|rkaPx zY_E}*-*~)Vm=$fm^aq3J>u>v6yW#Vi4xBTIiL>D5IVistlUc&eP{;wV?7Kye8DjH- zhaQBxszMmbqT`pQ&3xCb-Xj-~1WL^bEzck?e^hoytGTgxk)-eUSxw_{*o(33%PP?* zRu_Hq#-(l2>m7?e*q-{1*k+h|-5=&?=({vOS?WFb`idAZu~z(!y>ulOE}pcmjNRCF zy|g$%#f-uIEVmG{F&r3=BT1U2kY!BbWI9C~N2gxz`ThC!fLSgLzSexMK)<3>i|`nS zQ5&N?L`uA21G-hB4>5S(F%PBH8f*JH+UAY)-X*}%D;GL0%bV;7RL(=Q0wG$j;PI$& zFov?I>4Gu;YLnZcBSS0x9jsJz-`d~62|U^O4e07^q`|_2v4fBzW{n6bk?mjw>!8++ zt)R~ElY+S(?2EQ2PJuivP)NDUd=9%$+tP6!g3-b?MEY00cqUENYnYsC8 zpI@nFHcRm3=ca13Hh2NVI1RCqmB#!Zd+P7&N=8|GXv<($e?fs# z{tIHcVv*rgFnWvh?`2@P3Jk1-{)pk1oW~WuZ@$^8Z+mY2b+QJE%0Qvv+VQrx4a>8y zHyeV4)QlOo)g>E(kjKyua)YFV{5?A( zE&=@(6PDbH*(PwV&_fRu&FhQ4!azw;LLuRc1i()ug@*s!Y)4)RFm3aLDU=JS;l)vwZ>h!1b zkTq_qjTL8xIiy&d0V%wWQE^I7kAw&vBlv?@&-k2t$7DbHvaR`pNtmAqH$5Zh#;MWN zUVo>ROtpMvzImg))8Jss0oa|0i7Oqa40BHVgMJA_q!e8QQ5r#fS)Vl<(B|aKlq|e; zRuP1rJ6?6mg&KzE#2Qo}>Yf-5YdBwas~s*S3K+IM3`DHUso8F9r#gkiFEw6#$w~-i z_#ZnjwG_gwGuQy&F@*pa2}8G0qIB1^w&&BmN$!xkqiM{}6t;q=o<5kZvJ;}VnVl2S z;CNOaaKrEkFGpsmA8{wq_=A*xOpUl2gA68f2m)UYRH#jkn%~It?003@T7O zsFV-8pngnHeSYja137J7ouig)KGz|4EO8nKqj>{SF*}WR#z}<;3&7I5UImsp{mAP? zEMVqJG(SR$5 z%Q|(Q<6xqtDGCZDBAwFfPtKQrb%w)3F=`5cpZCGZ_Pg8;{=oa{5O4Y zAwp2ZjQ_s%gSUTWna<+g4T7b_LR0A>7qhFcmf{v{ZU2b7Jy)`8^pF5p#@G1alkVTx z;u*_eyI#38D#(NN;Zg>K&MUh}R#ptn?W*_VI+yF&87~bHRc;ha7_RPF`W8MKT6^bZ z!{Dh>PUH#n4r43qZ`v*5YaZPcJW{5dj{}jf5aUabNBxB&nKu=LkbejQx>&Q?)#tnW zE!WSDYqw~_c<4Kd5G-sL&VzJif$&sc9N_8uw3fi6c}a=RqO-oczw;@w`w`QQ^03$t>J}(kqwc*s-h% zhdlj#pHMF)Pv-}r*5z}n|9T}PBwVqv(?%vV1(*vI#32*C^Xf>y3)^06i8D`u!0R2k ze|vMlsN=t`CQnJbO=g!S+c$3Q;RQfpNUN9PMMX*B=WoJdG;IXOfo|W?+pFFNgQDpD zDJCT(uP*Aj0|_~z-MUw2Y~`m)B^#qO6_q$0E3!KpE(Q)E;ojcy@2g|c%81r`NTR8E2|rzt}JIze2O}qP8?*3#ZNI zYrs?r&gW!#a@J>`P@{SJerojm8?goy?v|YOD>kXgf1m%I5s9$B14$P=S5g>GM(B}2 zcqVLU!ZIT<1u2v`x8~$LJ+?9}cgS5w*C!lBEbIxAWBSaA$6i1;HrV&So?L}IKY#Hw zJ{QQ(^kJ|V5&~`7X85Iwr}nih?P*oxFZ|jN4tWHDkvA4&E2|7cM8psW7bCw?zJV*L zO4Mh?d}uW}H)Jy|X=%>R``<|peze~F|1$Re3#j$i3h)AsR}R8yxiRpFAe8H?ZfH5Ow*rJ*>uLR{i#S6nh;A)UX1Y= zQ2O{D)8^)R17r}0^mR8j@+g=8XmyP8hLk<_ReO}(uXJ|aMy0zr)qmwadArFz=T6bv z+l%~k%}I|*v&nmYfI6XZG}$0kg(lp#Y4pm!3mqvJ(c4bc>6Sh zZ2o)Vg)s=3*)C2B4BV~h{vsUwC$QhM{XPBW%{3^moYWRf&55_+uHoS}AeoDG8(3c$ zVyxN|bk#Zk-J`P>1;wd5wiGmjCMG88o0}__SiRahxOaeeb~^7=Wd#;z z;6L`qd=v3MhB9nGs4X-?J_l(r0acV%ws~j<6Df-cuq1%)7_KYyq^?c_{6|Z0j_#iL zV_$-%n?x0v17O`^tJ|1sYAWR<$StdpPK+j0mc6u7C-#F;KMfcMe%1hW8o{Xv9Td#zat|Zcb>_p z#2+4)weAl)!wnpn4Z8JF6UeFm_pK-ZFT14{VcrvSN$!4G2Yk!Q9xfm}X*2#&nmJRckr)-b}%iMLv+l)TfS7(y|r<9WTR%L8K6oR&l}HflSxH9yNCNf zEI<_usFAPumNcW@U9F_NPp6JY#6JDXw0!yAqCs}zE zT6eu{bursePz<3!RcuXGQ!QPW{hsj-%vyT781p|&8YYn6b9-hb|{k* z?W;fpBn01v0-UZF6pK#bbc%h}(sPR~J|m!U7k&4*A=si|9L0i~4zzF6 z82L<;9E7EpuRZFxm-jjD~HX`Bt@@r-<;Sf}l>0 z-rVPKT4leY8qpHs6QxIWDKD25MBEw&HceiSJDzCY7{5(BprD76>nYOhgq8cF@Ht(d zuVL(*T1RutG8b6x*yKuq0jxRfn$2gt3Q_*Rosbtk^o4PT4$9F1GEj$2p{VgY|({X~m=X;qq=U9Ue9$&f033=$4Yv z{_KDW!w@RW*M8F`i=ED>eNNr*)MPRNJRhQ+A8lN*(#~+TMl4-jAr4k zIe(RYG1H%0Tt~obC-^K`q)~O=&YyZl1HH0baS4=CbRw+S&S=$)GCV8BZA%Way}erz zV;pZ|S8#Uh4(x~hz09)os%6tioCHV<7dM^GvqQhXLNyncOPg82gql(J?3t~ND|=w# z$Brm_zHSHZ*3)X*a{3u~szSTIddXjt#chU+%YV$%`Cbv${#t`xohjjDIEzd6W$GEJ znlBc*+Vtpc0Xw9XZ?MfPh@}!(lAIU+#n&00{E3y1bl+)x%7oy{VafR~ z3iGIqkaTws^!Ma#m3B94+#xD-A#~B#U&8Kwg`XSa@BQ;XzRKmt*y#z)fpFwDG-{hMG@Zp`ZoZV~q^IV3@HUAP(=S*5>Ou^g$<^5g z55^U!r^7$+Uj4lr94VJHe{XWUZWoOPG?Eh?fipym;L8&%An|0 zZMl!~pz6|cp?L9do}r0q_)(S0L#x#uW@}{PMU!J~VnTmYldUR<)BhRSX*t@i-X>0? zck9p3J;bE$kr!;)q{>B}C9h#pMpEie6$E!FWcr5$Td zrQvbsr(Vs)dfQ>L>ECtwM?j@v7oEKP=3TEVI|hwrqw`a0robbcMNus88d{dcLF@WX zUC=LM$uhMA$sDkVOtu0bBT^&(HGt7>|9ryyZyL-Rnbw{JX%7_Ow0H zX~s-_8SrwPYl^UGVqbO#UQYW1wGU>>)SJX8=p9?REO80v<*!Y@<4mdND$ssW)8rdp zqmeK3iu&x1=R2SB&xNR^&c~p$1&nb)R5?O$t~}KyhllzeXHevC)so3~r;I>ltEff= z$x+&2TvO1ZS1tZ_9F`n+CinoNak4YL95QJK<_0P>0=5e#cpWd1+#tOgPZK(`g<~r;5LcNbgYS z(Yx$DdH?|cB#@NY!3~;f6ZJYtKNg_q6qkK;7#JZcH(PPxs-khdB$ucERwmEWVh1EG z6xgq^MBVffBh?9xz?4#F_Ty%_C1Bn9M+&FPy4=SFHpax2H@PgB|B;)gal3B|Z8)>% zCRlxsJ)QMB3^qo5%}iY!Som0Ok;SlSyG#+d!i0BL0F%b+QH<4IgDWXNL4 zu3oR0a(X(MDar*k-E7WDqU1-DvuFt}Hlf|&rPCs9RBcaj?PQ6%U^VflwF!f~B~e!C z=2j+tyvWEzHm7U!uNt492OiuiR@d@0;sRhdW^0Wv#eSZ$F4~Evb#Nao7seIbOmH}T zXFv~StA?n|bPCyBPI}ZsPobfAbot>A!(ZyyH<*w&PP$m0melr9eT=`3_ihZ}%M9 z7?QbBW;%UoE*wfzQ?@s0lA?2vG3e9Sm!b2}vE||Zh^{A-%M(^WZiQ&KedMp#2#7to zS;=SoaMtrCHyG#Z3C56PH^>d=*u2p8Z(h1sdbVsmg|-t8;m;hg-WHFoc#JpBU2H>G zaZ0||xIY}*UA%rR+}L^*^ymyifEGV`qP9K((gUQXXpZ-OIjGsADpwhuX|lYz{>+oYbzSAOM*UAyhyHu6 zmI@zuzSX`5z8OyFZhu##mHT+^QzI_cP;_97EY{gUd^WV=~Q;ejL;- zNZDM2IVPDI5h)7cY5I^U*NSV4hUNGZ1!_PSYkUuT5o=YbAAmSBxs6s>O$m zL%+)E5tIjSHngo7b}-d%HZneyqNpC9wL9Kq3e9=vNSwlN>F0$&&8Hm>B*D8Po(oWI zR-BOY{UaAtmMeXK7S>i0@}sKTf5uE_tyl)-63L$NtN~7hzmBjYJA2_&ZNQ-r>;Mo2 zaW>4pV@`AVgTeHdc5S)e4S@ktMpOZ#r-6;^?AnGPuNAUf5bh{7WP;`I{u#S?4T7}= z+x4-{j!*WZ>PUOR8?8dQGA{9gx@xvYV-q2hs8_8-&xEqJsd#bnl4z+o8B$H#Su^lj z=^U1K4T#ATRPaS?M?NeDz{}TjR;W>oAX(*tlGc8H#2eAA9VRd)iufQIeWEfWOgbvx zDsA0WB@2eJ%axF1>DZB;^L*j^0Uyfg5_R`(OC_|iK4xy4^ENR%Q;8%re4IAZ7wd{g ztryNMy4MuAhb6L0mG731R!Rj^O`@+kyjkwR+)U9E1mwi;pO=1}=DcMC0c)KzRJkyS z(`abazjw9U@M+bD$q=`0t}z>oyJ}muB-@HFtU=(Xe|8_pV8BiB35w}>v5=eWX5DYM%X-@ zlF*>$)m#0q(NhfXz2p~9`k^p<56A}(C)&F57mB#JNm`v034fp3M;{+;5Y(1Ce^>$O zd#+``Uv*q({@Bw7ARrLN2B2=V%y!`T5I}y5CRhMwjrU2;xY|RpFrO_ObzARr(b~jc z)flgN1A!jF$HO7?;ZOdvyhDq#10x>nAz6NiyDPL45A^{3J`t1azB$)u;n{d@&{=3b zi`iL+{|+#9J^sN^ap@d0pzE;7&{HQtiS{%w$e?T0cr1tU7=1kq==rNh zhj?$!Vp$w-`?)=y6e+HX1>R|YvY`_iKimjhpW+7x&L(@c<`XNHlY;q8hv z#5BN2d+Qd3y0TS)mEzoOD>={kabSX2L1BC2h{i!Yp;|Q0k{_4tkHb#ggDtk;J$Qcu z;M0q!fr|l>epRe2d&ekv_5~UW+@lq%`_ad5F}XZ#UUU@JAmkW|3t4CQHi8A9N5#gC z7eV5j$4;^B0dRZL&%_&@vb?qN!sbO!P{0JPq}XKY$xXV4oooq8 zU1?4&vol5rt9?EPQz*AL3?u(r_5r9a2SV@L(ZJZiOoF)i8e?jvY}qEOjz?X!VUlxk z2j#`U|FHJf$s0{mU=pwQSSF6Ma4DDdkJQ;B6RZ_fCCyo}KEy`*gT>S%bQ8`HYk4@! zlUvVNCq7~Hn5#3tRrtT_^XhPXUQl5$SR?+HcGhnFeo^7L-oR$^&GF6^JHx=?E;t!S4Uq7|3Z+m=_BS$GYm0oZ_9 z3c!g-dz6DxN`&r~cr`5oE5YWBhAL6L+y7w!Y-J9oyO6uzq|M>%_%km2%aRM*$SWU5 z-3}wtOCKD9bHezg=5iETk|hR=h+1nPdzT=(*KEm6!UzuLPLS2GwxdlFSK->u&1rlA z&R~546luYj#4c2b`c%jZjpEL~w;Rl&sd;|Nl7F?NzS#GQ2Km+wFMllxu{?144amB> z7aqlbAA2tWRCmHu=8lWqT`JBs<@|SAAAS-{1?iH|(EQSliAaMD^c$zJ?R(cnM^&xF z&4)b6#{wNffX373RZ;Gg=)jz4hlX$9N2#n1ouCRpoMRZVD=$Zg3Nl|{?zk!d?%vWmuI4&Nh-m97eR zqvs~?hthr5WDx)B=&xr9p6#*c`Ldj_1@d{mvFwC|Ox1=WZn5PN30!YY?BKQko~zAP zyS=!ZV?|1f^jVR0b_G;k2J-sIUbVW80vHi}YKf(x?rpf|@M#uQhn?am3Qew;bO_XY zuM-A3ddXJdHD?_T{+!Pj@M!w#pwJaGg0RGpK+A|nI(9xj-_~ZJ_AbjBa*#`p)POK=W z*^p^hV@U!>? zL&dL!JhvuO6Kxgq*$P@aF8uCnyP}M8tu}V^<8js(O|hI3sRwMJPEmTUckqAqWO#HK zLro9ll}@l9`X|eeQx~v4olP2AxLMN{x#6iX_vcEE1FeZO^-lJC0!U4Ne$Ia|5T>mP z2_27>3c-|wtgNjcvTa19$I^jV5a~n~kCioyTX;e8EPyNO;vf-Jq&|cLs_*W}>oc&A zKu2EeyBgVB@p!0TLM~=<@+C-Ce8gqr#^jC=eCRiBDg#=klK=5rAo z_a&Y%wqHAlDHlxL2fm^BN_>=W2YFXUf3?G_XbcZ|o>sQ@ekEwOxMPgieAM9lcy2G( za5MnY`zS}$ZKW47OEd??h*0(!R?yF=Z%=~5@arJ=?xZ$nLsTcwF!o-VdkxnHB$u`q z@IOzRFNfLjyO)o&FKeEz*DeKM(`yCrhxP+Ce4;y$FWX{c1sK+vR#4mlnA=`3EMAX)K2%!2K7Iqa-)b^pEsA-jC!*8 z0Q9N1hZk@0&&lUXiM{s{xmngTTjM~xJ;vL zXqE7IX04l46}b~6OzKTc8doV>A>q+XfwTCve#ican!M$U&yQM>2I{I7)xRn8W_QHJ zI6IpY=~*MU7^!MW&<{KyXHsz@V?hN@PJyNnVbHc{^P{!H6!J|X2SZ5ovrJe2j>6X% zFC{zz63;{}Brr`FWOTF_ir~@{r=orBBqhBfHvRgH~RDDnW95UhD**RM&D^K1f_cXWj~;jZ?wsZD&AXX;y{ag>9Ts;pb2 zZv~L=%*dgvT>_2=5c$*Cxe$f6`?Dq**0x8>^Tp3^!hu@sU)F1i!=cN`M0#CWqMxM| z2nogv29T(jOyXFcQ&&+On`liZoZ>w5)D+U=T8Q8O>ax~UF35?ENB^Mpgkay>-@4Re zgD)%m&ui$*dO3AZSUO)FFk1j^ERNqop$8L#X65&fM~QH6f4H-rM=&0xjHL4I_m)0L zBqro*)Lr%^D|B|f3)YV|B$(!9Vjq>yH2W$d2TYOV{3I11mxh{cx1tPWw-U@w2;$8 zC5o;cNdEkyj{sUAnz1%o!8(LD{ngU=13?q{uR0^>Sma z1Q+IR<{16JGgNIlcG|@xe!^?Je!F+E#8X>YS4E@z14E79VGW^@hh^j@y~X3;7^WK~ zRQjU-MMhW}`K%JX2cdkW@(nYblc)csyp zH~xE#T#$Ufzq3;D4`>A8=79B@b2wF{+mZYG4UV?O{nqUy5DyjQqP}+4RS93y;9DSi?-TZ#Uak1LW)7ofqd{JRbNnqa=_jiQAPxIgLrirIi zv@RYm`B6c5&@t|qvbXYZ`bcI$`3&P5%@9$(@SQBScExt)eCJ47wQ?|aBRCS#T!fwB1_v!v9ca9gj+1l`#DK%uk zXSz*|>WSGI89&YGa0fPS8_6XGxN0%T$@rmvvu6+4!$`F8+0s#B&hKnz?Oq;N8IqX+ zzfLb%7`mb<9z=*f6Y{>y5pQ+Eft%R=w3_%Ucve((3-y7F1iCMf!WD3n#4UzRa?ln&(2P^`Xc}a z3>_U+HWn;rhBMkDo|g+1nuZ%AB$duZArsiY@#GMPW=ZbRhPlYiMP^^kcb$<4vBII} zDHB`S&-+=9kL_GM@zXmanTW~~^jh#(%8_u`ncp5 zjw06PV>GU9Z#m(}4R1H8$0cfv1=oCD>RH!w%H_JFt|!B&HRwHkblbny>TBw-Z>V0g z-OJm`L@3@~w)W^FHlZT6t?W&BX4$0*73EgE&zDZ~XoJpY8(6d1%z}sCBV-5E3_EJR zhmRBV?17yA>loQelZxUy}4-&!J{B904_Pc}N8Fuy{| z8#u=;qv|RdIFB+~X`Oahn7oL;D}&k;h*@(gs#)7{TBm9AmeExd8F(={sJH7}CG6!# zNgA>K)N*ulD$=TZXNq$;0Y_K7tgsY@0liRTOwcZqsq^xbf;!{OR))O@u=|LK+2)xD zTPq+AAQ+<9cyx8S!70R8ECMn@@oY+*`RXVHid~;4Hw~QJH%!kh1yj?2Xy0`)iJLOa z(h|}A(oZjoT;tWej(mYUS#TO)4}wJFE~pc1fBXng7z2~j*|_+O^F0;AiZKAVQUvA3 zBt`hdM;D4G^$@ei{32mWK_vw4%RYwInVhv!IY6MUMrZm8;?tC*s4-Mzb)hHP4{Qz3Zb`U_q!(pui;LCdL6>HMc_X**QKaht&bh7%09;(Z~z; z6w2JQ9yiQvjgE8{Hw{Yi{TWuIUVML3a2!$;D91ZNtVR^R;IaKk!UQ?2WwPW1>G0Q+ z3M#(fQbD@|<>S5mX3Q!$Mo}zepuaONTqZLKC1Oj%e$v!a{j^;s$TkQ+dXGW2;)4gv za^!l7s6Wmd74n8PTiuEf$Xi4^*HWh&XnN;-0uoEB9+m=+)a1L!n!skBPfEkB2t?&Cu%A&gbZjC?cQqSn;y`S+PXIgr% zY+JGswRtm|i@i>KH%4+Ia~r4G=nD@^1EmH-QYgnHd}G1QgY9 z25BN~iuF-1F2kl=89D(Oj(YcdQo!6;#v>_JgEVy7jK|ctO4kGlitN-&WG#@w`E1;MjErWiR}+Xax!y#09kt+r{9a{+lbzgxW7b<)7%Q8h?h8B+mC zCp6ebVQrkb{X6J-i}?xb<{qeT;nE%FpT@FBX4+rtO)yLGjG!5Pynv_dN6~LGD)xh zBoR5=U7O!Aq55Habr(jl?{~#pJiNmTA>r?ATJtU=TRs>hNzT+vOeMQ&jA&lXTlH!+ zkpwKY?kDyaO0b8zoZNJ^YP;L;KqK0gdbY93;b>&C%`Lu$pxoVRS)#x~st%R{lcC>qk5;#b(ffIl`30^>7UFazR&_fnTsbe&3(>BV6!8$52dRdG?`l}S*?wMBD z(d_)Zeyh99)-Ckt%bN)b_oKYvKC@%f)2%3Jy?O42sw$vzdI<`-V>7n#pBo7xLkob+ zcS=bQ0gW#T$4fI?C52MI>U+n-XSS6A7p-OPm`HG)skD_;t1>k*| zfO^&=B{aXuu0%$u76AkzW4_w)3)B#I%>#&p?TmEO@Q}o z{wUh|{H%K?GF$#}k{U=?SKsdRE=vzd9LPWaeVA8h+I7q;Bv8}_zIkDfY`)iKe*R<^ ziP*0PZAG< z#C;Ant*b&;jk#v zU51c0*KT*7-pKS-btdGb|EW80dl?;o6Za)?RKhe7gk*nRsPG9rEpxT_I_ZnR>);47 zOQLu4hwYP+TNQY)#8)+y=@h<%dX2ti+6h0chCQu0qy#R{ew+gmA-s!#D;=!=Ds?{UW|*sJ@1VV%%t`NDw!Er~&@br_%rq^O>? zARy^xT!j5(MkG>(GO4|)y+zP9JSP-f!%Y|Ose!|k$!uBn82dppBa#sZboa`xndFgS zv=yxa4O*X``66tFHbqW{5gg}+Z9PxlBY}ixwp|)dy#txjKF5a3({YdETzX8p)1NKV zcSBVHLtobU+IEskm$I*MBfJ1UJ2Aj+u{LA*_dI2~CrovUI;Z=5^^?AF6Ua&ko;|}I zs)T&SWm}tAkztlsp%pv+k*c(GVRA|7$P-S+S#=p3^@VE&L# zh{?~bg_??hIjB&ptP`^T2}X}ovHY3f8@gwslEWYZF3bOFFoHh?*^dGT5VoFgNahe`kk)o&_3zGX;jA%kATZ1reeNoiUz8>(R!gDLtf?lcpSF z^Ru_>03eOK&i!I;N*JUk5D8k9sR;Oj!IO-74`>bM?k-%yQc@h-&TJ@x%BusreA&37X|Jc9cLqe z&;U@Y_+H9YQ5Iw`Zkk*2TRti09qap=++0s0F-;vl*`)k3I!bDUVD?^JZk(`GAcG*O zVxINea4_omIOZjokW&1`upzH z$8f~LyS-YcMWVuJg13lWT8ekScGz;Os@J5|1oo(OIWC*C2B+-={SG30@hho7>gQmr zK3S+P_wD=mZgFA(M*z;g!pKx7b%av;c-RG|oeB}|KrhByOk8aj`Pv)khJ(|jY4%@R z-1yWc396g!0^G{A9+yhEk_u33a6AQ5BOOMav2u7VW(8xJAP&y;I)Nm}4~Nf0TxYkY z+fARcq;MD8@lg@}Da4td)0jD1n~}B7+w6Z|1SiIs#)vzXDT6uu{y|L#kJ|2gGFlp3 z0>r`?rhv`ukFjz>5#$EGogR40#0qmE-pF4784wEzS;*EeE7qIi&184DK(pfic@2te zhr+*FUUnW~BA$Et+;F?Z=C)Gy-dr~g*N)I%+|}6LH003Ar(oHgogrE8_9m4rGM*XA zR^sqdGBYQ0W&svZm_AF|?L6SKhXhnGM5;4uTsCBisMPOHw=Ss!*o{@5&igeF134Hd z|H}^{i3+=ZhR;RH?@4Xcx3^t=K&2#}W9Q&TR((o6=G6Tn%CZ8w+-EbDDS9VlpCZ&N zUQn+Pig5sNkd0n2B?QEnkDl1av3ClW#A<6ZXDB&sKYtIbkg|SmecBcw&Fu-$0txYE zL?<%v<(3xK6PfTcrq1#Ux@oce*ga1t+KJuUU%$N`r@am%?_61>qS=2+lIh3yx-=@F zAd0NYaYsNZpL#zGEfC2ysUeVge{GG6%FTV1!r=_`Xp`nG2n}dTc#MoL7tGGZz5F+? z!`GCM3vVvF3f6B^0{gK^Q9}dxum*c`113DWZd0lpcb8M%J{4%No6443?CyptOL{kx zD}Fg4{t?;t(ENrUC)o%<6{zhJw$5V>psZFC4`pw$)dZ0r3G$oics%tcWR{>b&edS0 zesz-7^KKC(tfq}4IEr9QBtDQ=yayoubiUc6^=f>m74|M5h-hTP!c!Brxx10ka z(0~jLvVJ#&2ZO%$AC!@j3E(kq$t~MT`v`eD#(abp3MV5WLD_b~;TtDeCU`&_gW=xD zlFvI&V(+B{A1ExmfJQ=zOI{`HEcXG4GYc`WN~`;Y~PXyd5<^?8j0+=|J-(!5|kwpj({oW9m{}C6Yob_6gde)k37w;53MMfHzQI-gbLb@3N_u7tK+){TOGer z((>@|^P4?-m_A!aZI@%8rgcLYV`iK}!32Q7N7#k+RaI5Pwz7yV?B`lf+ZbMXW{)hh zbpb_&i<|BCZVp%|;I#1Gwl?yI%1T{)KLbhJ_-gb0d_BOH6m4@MWb6ed8lrtF!*8h; zviF*SSi9KDIQPza;at6n4YeXQt-gQE;bfIdQhGWd%=gmUgdxbyPscd9Gc*iezOz7Z@oAk_w!jK_}2kX2{IjL4?~2k zH7>?_@-*Piq`bcNpb?HA;Z-G+M)_ut3gr@wKjo;`OlV(qKIFRXUm^14F5s^D5FLrnbISbG@D=99R|&1*{x#iykCZ-ij)Raj z2*AN!=5zcU>z~-P62n&k+1Msi&#cu;4*$3sW`AxS9j(n?dUys@lgqqgw62Hh$!MDE zR0|4^nbxEANof;nIf)tUrz$Zb&f?@+gce(0r9{p>cw|)bG}Lbvi$^+7?}mV`(z6I{ zc>!&M-M20oOMv&{$x~z`)7{(BjUUjRryhV-vgrtPn(bMLrWdRCK5Ty$b-qR$O;WpM zl23)ng)p#*p||8($!p8};Ygr7#{0^qsxueE{hf0^nBz)Gru5z)E(e! zx+I0O*Fqs7@|*H!hqFt`>FGZ|z7bOpqE9^@2$0};0O<Z5dSsq1OM%}{lB-dDh{|>>`iq1xLieba4Y;7+Ud4N6cAQGxdlH! z^o@#?S7-GoDPhNsHJvOkvlqhk=yJMEX~-)h1k{RuS`TE=9vXm&b|@mF5V$ z*l>E@jv<(G&y|{ePd)v;E1p+ybb^|Mv9d8vs1LjP_O@UP%NNR z(0DEd|je%+S$61f9>Rix`(G2MPk7pu+cSKEK5U|5Z$Q%`KedCOi}U0D?qFC z6GJfmFBjlUz3^#bAg&Buhie01?*BMDvXm>HVvjoEU_)Y_VB>A=n$kh+d;i|+l5xTQ z1xaQH#uF*%dUp?|{(W$;NL#zowDcbt!8LlIzmOZBH3`hFPfF^q-G8_ z`x}K`t*+j!m$bbj2bH+D_?4sa@HfQ=g-5z*=NL^}bq zHXhzkOk|TLs0T0lU>>_%Z)hFMH?Z0ToL|P{^hCY@Z$<8#KELbkUb^mH8j;Abs(RI; z$$>EOFtYHle$~e0FtUYLIB^rF^;5&h>@&~1kLHX&QP_=}15_WBKg(zpr`3JRZYCzv zq3zsrJJbQtOYV=!9~T@9O^eDjMUB+Z?(fSH^#`U~xHyo*kEM$8iV~~jT401F=OvgH zf#<}+tcE|OKC6PPFtyar{FK{|=Becf9?&~G&v7{YAxJ%HW7pYD$TV%6!h3RNt;QsO ze}eZ3d1eyEIIzoEZcDhw@TGRt010Pb5O+Go2xw;I9&AVLPR(keZBZN>*d7N<0arKQ z9F{&a-48Y)a$ih$Y392pgp#D%-( zC=#Uc7VwsUQ!G9q9t(sN2wHI;obCMCVb#!r$!h+xmyTKv5b6+|=1_cCn5mI`h{N?o z5Q%atkHYeprd~p~Sah05lf~~}WWpn{AC?v2+?5b`YzV%zRVpXn+^lm@h(_RB#GRwj zcoO6fKRc_(>$f0 zT7H0od-valRxJx8#}1yVB%wjnU=GpsHI{;!=Mp;!=KZ-whQUN=L^80vbtVF49U63k zaw=4P*pcC3912_(jxO~z)uyXwldDxzXbL429N&6oi>%f!H5;9W?s}u=?S;i%=RN|Wi^w}<|W_Y>45GjlH7h37+@- zlZc2E&xY7~XJBOzPL6&tS_?9WAzSd>A6uom;(2vx;@VW5m;3eU z@qWR&%pG}RiF#3`%V0IdA$ny)!+fRpmxbhjl?u1c!pPe1beX3sG>8xPFU)1{tZ+|X8uC6}ar!ozuW+lN1?S=6kx?$DQ=K-GL|&Y)S~J3DsncCnUx{)G8_<3vJeZQ7tM zwqg8}zTLI!{#dy}XxmUKH=}an8_z{Rk+4pP_`N%&UFTkN|7Z~T`DeU87X?b@yfz?6 z`k+Zpmih=6J5LN{$4E@m)HN+PxuS1$U`ajOwi6$nUnA6=WrMzJr)CS2jK4^|q$&#M zHfY(u`&~gnGl^NF&*7KYoPzEjQD!h`8qW?d^Vj5y<}i*&hzN%>b)w|Ykq0#c)`B3t zbS7QB&!)>4pA1q^WPvr(OudPNHxm`?!-o%Vx`E;3c>$U79JB3At;y7KWG59D?u>ew zZkEcG`W~(|;qE)!Ln)^+Rx~14#0sQERu2+-R0>@sRT%ab+yZ!(cy0#7up_rXa-!g*HFV1b4 zOZ6JuX4^TdUYF|@Rjn83es)$sL-Ybu2L|H(*PKBSqdvHlnG7MQhlYC58!PSG;=p3? z`w1ziM{tldl5nIl9i5I=vF}OKOjVzh`#MU0lW$Awk%O*o3vg{6T`wRL_Sf#1Ml$of zt~0*iSo@7v!V?y}vSQ(YNvhwbyK~N`FdmX<=d-%3oEOo~TUgPMR1jL=<16uw{%(G5 zZrgX^uGjZY{Vqd*t=@il>$-<2ZZv;Lll;;Zh(XH8NP}>ovAJ1apuyoPrgh%^phPt% z*%Ers?A8kE+Xf$^_i$b7Yq*(+vO=iuPmf*QYa3!gY#Z0V-B|D?7A0IR&YYZh502!9 z-NNt_MX0l@V`cqRJu))7O55$xsL3Ou%GX&ES`L}MYdv^Oy!}i+7Tg|)8aDwZUrm7$ zRv_(1`uH0W%xA(BYyoFo9~lCoAw{YeEkf9x9ZC`sol#nMQCEJ&$TA=@g{Xg8D}UkN zV6nIl+W3DJ?A< ziQq*q`=w_(Rq@&WYP*cpPo)a!1gf(r_2-I{djD0<%30;Hjk6D>x&cI zq>6>7q*;RgzZaX&XjICyn=iU|Tnv0K#k&w4o-nTjhQZ@-UdMk4(~Ej0Dt-S7yEuAO zdAQZ1qD0DcQwZ?fP6ZnJ=-61u>RfB;u&;?pRS=NvE2cN8g-t zA3;J6+euG%o)1l=>IpQcArh4`C!>ImVYQT*(dy>+?Ojtc5+XwEXJoiZZ!+I}ctLFQ zk`76s@e52j@1jbX{g;?jQ@F zfD8!{(V)@scWQ+Q{!8qlKap#FSok~nhlV%SEVe^%ire;{Ne_o#weEZ!_Lm~@N4%WfW zPMgt?#D$V|9g!0fv+aDEMV`iy|U!Nz9&`*&~4 zLxBq={0h_0`yhBaSn_#Y`+|bNqJXo{gI?65_iOO^*+Ns~nlvV61F^>Yq2z}10t4Fn zupG**p2!Hop=(2+Z#i?`-A(;MQQ2k}9>UxIHZF|5LF^&i8Q06#MWC*6BnsXxBtiGs zeuSr%{aLS(15TMs^>bog-lapDjQN|O{#U21OYkBs0e7+Ib#7N8F8Oy$ZF(-FW8!0q zLdQEvqQJCyv;NstnI8YQMLo_BlE%hspQGnX)XHCfc14jRF6)pCz19lBjfV=T1-0Z4 zcb*vF?HliwP9q9R)3$+4z(A z`aGLxeO+BFJ*Fh$`R!iTe@8=yM<)^iq>7eFm29COU0>jTdU@S1uRw!xgZj$(L5UIG zU{Wl-Qn6YNY&tW5`tM~x5VoWjD3|E``R^al!;rl#esf?}d`3h0Ln z;rdqp#T1;)hdq&avkiV`;cL?Oi)?T|(_!@Xk=bV@NqTP)hYo0qCzXc-84u7)zAIh& zJqbnhjn>lA@;m8~z49n$R2|XyvlFFvD>bcE#614&6RXO<8p20~jongi{+O~L0A?tN z2#dP3fu%K}Cdx|7zWnw1=;zO;GZk0!@ZKEVe}q${DoVhBp`(DPt1}YF>g*&8W{63e zs`~mG$2Ft^J%>8zy4`Y}!wm>pZX7_>9WhiWGCVwi{>;&_^6c~{^XrZR_n~4?&#-=7 zk@&ZVe`!a?C?F$8*H=}$I@bbOW6%duJ=7rc;r)3=P|&nX1ov-mxAu1CtV?#9fGh^d zeodCUsE;>mVeO4nob$$0^uNcFYD&@8Sx_&kwulP3?@G%v9=Y_B{PcfGF4$(b;fL8L z7XA5RfeG@Q^U}T&asPIdGRMm-1bY-$TVxOXxKL$cN(c^?)|J2^=DttZfR>E?c&>0- zHj7x=fkDvpTfw_J_fV6y`}OLd#u2|g^N68tB%P(2eZ-IQ~y>rSi z+F1h2%BgzGFeSrflz+Rqgoh87u-qqB)Dw$0Oc_`+TK^=rz!3W?CPmoY5vI&DP%jn; zB{D#HACdHD?19lY*qI~k5@Nf@OJU~l=}WJS;44X830s2VjXcfwFxE#IiyqO2s3V3mF?F-aWaX$*}p)QL5pS~lC82hryu36boQ{Df4 z3XN!%%<@Amm71njv7;rNM9$w=9^^{2J8I~k6hcaeQHBU}NpLJ5SRfH6z<^v01O0u{ z>u;m_g!L@xLl7Ky1Q-(^RN&C@i;!qIXeopQ^#9u=urp^35>_v^*_xwJ|NbGvAj4b4 z%34)^D-(2zRra*~Kf-9B10Fm{lceO5P6?z7-Lh<^Vxulbs=naV;wIl-TARAQnyUhwmeJgBM)A9YH>M|bB@7deQ*2XzdE zzsJqiVC2{GzGRk8?noFM|F)?CrdW zC^=7j!gwe&>ai!C53$W&8uw*ROd6@>8|pd%N=E%M#op@ZguKv5P_Gz{Uifd&M4Dp^ zI$q}xYKrAmJG$4_sc!z7x@{sK0U)R_kZM~Qp|#0Ik83Tf3$017Q&S`v_}BIHdpJ0g z6xdblbH|UEnYkn%&&+CXMNl9(%YG7nO-I+)P1~C{>=mA4>tM54n2w?Hug73FIPnt@ zb`*B-9v1~0=sfY?FI8=_nCV=2Vu=+Ohh_f)YhUJ(4O0U6R0T%Yl_|s=NiUcPP z3L*RVG!}a1xX~WZ8rH|BBnjiA!gFq7kU!*fa2f-H1|TIru62EgL3uX(?~&Y zn5kc)COp2=S_w!AU{OiRo9?mUR?7WOEr+j8QuuH5(xp9QA>f20fb|2GC9 z!N1-%<@tZo8`oqH9^Bp4Azu(Zc<=zMYQ + + + + + + + + + + + SelectPilot + Deterministic local transformation layer for selected text + + Turn selected text into + structured local output. + Action briefs, rewrites, summaries, and extraction artifacts + through local Ollama, not cloud inference on the core path. + + + Local-first + + + + + ENVIRONMENT TRUTH + + SELECTION + + Turn the selected text into + a practical plan. + + STRUCTURED RESULT + + { + "summary": "Create an action-ready brief", + "actions": ["Ship MVP", "Record demo"] + + diff --git a/assets/marketing/selectpilot-small-promo.png b/assets/marketing/selectpilot-small-promo.png new file mode 100644 index 0000000000000000000000000000000000000000..0562eb6cea889c651cb1aa2aaed47c9c68553f6b GIT binary patch literal 26747 zcmcF~V{jzj_xC2**tW3~Yh&BCZQI(|_68GdY89)PX}TL7Mq6b zqklO*7RP79!^5nnf>_wZw26BTy1u@;dj?^`K*f3_Jv~A_Lc&`O1q}&`nPCGe!V^oX z6vZK&x)ezY5uhSq4KYQTVrE{Gd7dt%4o%B6CgY4eRCqK_kUCZj94N>r=Tj^M7*y6W zZPb*eOcDl~x@fljM1grwp-nGCla8cALW}G{2o(mt)59D%Q%i>GmSc~VYD$^Lgn$N5 z=|QzDPL#~|##a-LRNI|3=A^`h3l)Cx5&Av6pjcX%Z2bF&7%*jUFjQ0$Ks7hM^%et)#^}%$&*Fpbi~n`aW~vCh>Mdel{&CnBP%YJC=9fe<}J(+Bo*%)EAA%3^7GT9 zLXSvaPlXkxp89RYXj%!R3Bw#+#DRH;6$RH#>aMQt&aD%~ng(|0?qNnRncND~8d|YZ2rSk3UilC)6 z1rr*qn_C&3G9V=ej2j)7HZCSjlmezQ21NRaD?DmSP^&N%gu?$$EN8z&V37}8iqfVpWC3n;Fn(LHjOz0m&FXHPvXUvC#$ z@nK1lty)Ky^$g{b|3-8UdG&6qk|UK)_;cbqY-lL=r{$bJtu{DAam37wW@1JX9kdWE zV-)kDm4v~-F=^9SP6L|&nXZ51Orp#e*M>GmMg;$Y9vQXtO{t!lLzI+8JV~C|q&r&el=DEGpe68Jh}&E2_cdePX>B)&Z6SXUK5%D=!mQd)JiQ{qMd zs^yP1f-*(Wg0KxrjKHMjL}0`JQ5302m642l-^T@dX;b$0cJ6`NWJ`>ik@ij zCI%ek8wuJ7>$W9DN|vml&=h&T5dml$_>p+I?XjJC+mNLk$a=S{0;n;hn}a0N%)Q#! z>H&wsWnaTT;DpojCMhvWjG&x3Rh>%iUENL+sNfEXLtKCDfBjY*n=bm2ppyoec$&(3 zn%Z(3EUNxhq+nu2&BVbynf0~680!%dUe+OJmAR90XXPm_&?JLLh&(`pGNVh-qw5Yr zr4lE@hc|h}q6)oTE~F(czB{HHn}9MZ3)~s09l@y=301j+FjofT~odIU$QclKxW%v@DdOmSD|H8&}2!QX=^8XXYkgXwr61DV;U5hX$m9 zgF-;1XB@1cEk9X%9T$_Vi*Qh^5a|eE;iMZA;7m%PmqoOrk7Z@NHYSau`>RZ>JV45h79K@zy%W+xZ@m!s?8j9>y|dU3b{3_-dj)LI~eK(c*7L4knZ z=i`2fg`lA!P#AJ7h0ff}thb0TQ}*%n&LiwW>;!}Qx51M_tl`7s4LmHYN@-&)(|GE< zJW!b?m59^#b}>8N&&Q2MuO7#``^zyt(g=Nri}Qe;eJn6r3TR5o?N6F$wfTcHMy*l! zQC$`a6|ASo0MK7vlv~)PrjEEO^!qqm10~a7V2eTTb1Wlcr3*T4+Mtt)j3vt8>LkJ9 zD8OXB*?z1DlGxx2?eIMLcDdz=j(##zpl`&96^Vd9wR3A=rBtXZEBZ+85M%zZZ;>Bj zI6r|$BaUOb8mQQoHoxp3pu}10qjyL3fD0!E|9nde6F@)DA8o&+lj7v$#1B@v`fEmR z>-Q0LS&=5)%k`;^rDeuk6D0gWze38}$>D{JfAJ_Y2M4WsQNChnkpXgmWiA1hK<6)1 zVu|E~q^i|_(oP~0Umin)Y%k@+m$@I|F<wZC@v=br3pZX$(^%cKY>WGfZ8p$qcRfd29}laK+LX>rHNuB0x(pEyffu zQl$em;)oQ*{7$|%7PAe7&hM7Vf7jG%qxX zirZUL6Z0IZp!|=47i)=@?CyTWGgd}%&>>aWP8p5ufBSDSRAxCMsLk_bhzjvMjaDaV z7RlITaKlCHkFB{1Z@4xWn-D1--FxVIHa6pSJ^>1fYL8V3HnQFw*Gh}^tmR5RKgtH_ z6kp~&DjJ$Jo(TyWGboBA3<^h%z&gp1E&W#3z#bZpsGDCTg4@Hvz#vPMR;!?hh>61E zHI9mmtW#-#8ePu9iJXu_6z(_KVHZn~fN(f1rfrb~`a1PUw`m= z`$YFq!mhTCS1d2qPK9GFi96!YJBhL?tQ&XM!fuCcOIOX;Aeb_L{Z0bP;ipsAJZ03pxK%FQ} zzXTQiU)4^ZFIYGp58})1j_5?Cq~lh*ia(mq_~ifYwks)5s$AGmVCZBt>|MF)I?NxW zsdEl@_m4rlSLiFfBFN%#?l8M52hU-%HpQjSxFJa7hcV_$#D$ZLQ#<*vp1Xq%q{l&9 zw8jf`o9_erC%c1oypb}7v|_ob5?M-aZslI#yx$SQxg9{OQ&L*^c6WOR1Xn~!V%+YHfuE+ z*e80K%7b)tDL5tgEtk8hXklujXkw&*1Ct6Qzes4N7yVjFo+q5X&;?Yb7f_VAM|w50 zXt6#0^`+|Ie)kx>1tFft-(DzNw#Q`BwyxGFDFl%?GbcRBM8)VJpzNr)#9NGbx!}jt zL&;pX7oXQXhJ3i*=r(J1E`&-4}`QLT5Mq_O4@fe9u4 zzWgom8?z%}h$~qkr4{BFl3kAc@9q&X-n6Q<#u!AFI=qESr>@-1HF9?0=KIH+>>f{g zYK0t5--TK*vP6`=+2Uixy|ImYKY?7tQH46=Np>>s)(O^O_q{;nGS}^1^!<_vk;6wl z=;1hm4BC>-dNmzRJMI8$hlg^^+SRg#^qm5+$%~HonNxm$$MKe{%T()Tl59QBCXI`0 za^vJA=hLLYOx3?l-;sk&pj%X{baNA-TYVn!#V+T7lLjQ ziY~`nM39ldF@&LZUvIcA^qk)-jQTJVhi^VR8C(l;z52WN3 zH%JvKgkb5U$sH`)*`K{!K0I|`ijAps4S-QyepzEOc=`dsQevuUhsZyF+$P8=kkd#+cV8zvTCVv0 z-wz<$cQ@7$b=5e*GjOl0tQv0+t2F#|!Y*zNazaM#aW1&Oy#&fU8jNUgj>KWvsuead|n2 zgoT7@Wn_Z_{9~e-Ftyx-XmwMApmWWkWHC#Z>h#@^_SmgGTa4o83d!3BmjuFTHc!ds zU(s>YS>N%2Yx!b0Gwq*hpSgL(#gJtC?|3dQD%COkO^W#k^o4taQJR_*k@A|F-*M6M zaV_wjLg5cQCdPZX7fUT+ztm`(9KI_7w8YY2Dbt_D-D;X0f$#6}VsO}o zb3$2hLx`g(#gumIUC;_T9{+|%Hsgpl>~DslzVxPf5%Hl)c&s#QZA#k5Z;;yf2F1U;JpAKuS424J`zzDrc+;=G2 z#)Fq=jiSJBEevuR60_=!A9+--F?QPQq=(H_*)*;j^^ca)Wy)m)xcsqO_XA&8A)ykE z6~e;8=~Zj8V-f9tYs=M(!CxXYhe08zm0quAqkc>eP9~AXEooX8S;%T3pPib~peq!8 zU_}$@!f^P1T!7OZd)G*u4>HlPx$cbl>6jcR4|Ru+XKc>Tr9oS)APjz==9*u@$O2aOMMaMU^jx** zJRY`PF4IOS$b@{c#lCIIW-{rHJ|A%iIAjUAbnT4~XR1A3Z(w6?_#UDH+<79o5(@gx z!4BT8b~OuoL>yyWq}3ODcmxD1`^UY<3Yng%SuP)zdZG{_K`Z#Xif; zTl}BH@~sJWI~EkuIsFWy!CV%-<)8pn$ZxWl)zegDbw9w>`#xE&&LH44VOU%em>`$U>)e~< z7E_;a-<@MKT0xgeo0h|2(-*^K`yIAK@cU`YqWo*#;0Z z*$L2SIvTZIjZw1i+GYz1@}rS~_SY)F}c(278!U+%7kX`wX98L7IwxC8*As z*qFE4%wFn4=kxiC@Tb56_Ac!z3>TdVWhKIStB;Wo@MeO0D904n(Zie3NcTNrx!#=L zPkK9=s}F`K&?+VA2#n-9;Xl6Y3wt7SX^v*Q$T3CB`X;Q?vWGNs_ z>oYRM`dbAAVberHL5!2)g5dw_aGf4gsG6Y=Bwr&4RAfo$87Qz_sC^ur$9^FOB~nbu zqPbR1aV+S){mh>e{=!PpNQH$J6@?|^-Ekx%N90CdXh7Yc9&p{TpR_Rup-JEIJN19^lZ;M&e zbXeQ4TF>89LtoB^cjy4wvjW1Bks)$6bZBSUc#+C^+ZjnjhPA`eMM}(j5v87qg4+`V zfuriC5LQ>Lj<%7vh_(5Di?U<*@F?+7i7@vGQT`Ny_zHVdBWp_lnqum^$&UJdm`)a+ zb~ZjU_l^E15ClbKABEOK=av}*W#(dq6%}EUkysurQ~rS?NPX>Gb87j@+o+(6aGU60 znHU4cjLgWAN8f|VGi-&^V2kbdLVh~Y93y|-zu+6N!~38k*DnvWjGknDwrg1w1^U%hM+hGHXYwQjtSeO`eG;MT^qFgODpJ z^^AhurRNwigaKT4)=n+JS1!gq*$gc|-SUrAlLi1C?~eri*gsg=)chGq{OE}+#<(QX z7457Yn>ptWmhli}34RiYP@KAuS!v_Je;m-MvNp;Nm;(b*d|e-uW6HhxbHj|1SS7{r zh?oOWlwd-@vTglAxe>vgbOe~lf%pWVCEs%v_#FTCejWvFI5aK{BxtfnESW#$Eb?*8 zU@VnQQ*a<5l-A4(5Ree_>K$#;mmcsJ-tB0vY9mS~-==&&dd*vh&EiTML0__^6Cx7wu;?LjcJwk*2&^!&Ukha+k{zvE)(CEY>5L6b?`3csMhWN2zD zUy;kna#=X;apk5=M;MFMggn?^PdkMnEhY2zbDqN|t__R!fN7}>T_aL2x95}7VecD0 zEqmMLjY~`QrcblgEN6;HKe-1{_w{Zx4|(3YuW8TYN;(gcM+o?fvjWkQ+=CnrhlL2J z$frJ@GUBUL(9yAhcXt2jDOt00D1%95Gi~7G_;~x#EN8>6u#vv*_j;o5<AWScY@3gZ*_c|YG+LYmxA!&YN5JQ8D8o+rKkvI6+-0AoG;_7_^uOiv>6gNkN(?Hz zUrj!YB;UC%=qgfZ<=uU+V{(FUj9(x7pI9j2s0bMJ)xO8bIWCnG!D(Fa2D3LgOm&xt zxkoUXfs~ z0YBJNUY8u2bn40ZDO;r)YN*Cl52?EN7(sh`I~#4fc<;grO4_o9M-5u&s^~gu;%x1M zwFbSdU7@CU}2*@UoL6KUjCue-3&8s?>pSRs*z}V5O*Cp#LIhU z?m^D%hU(s)SEMjPK_9NHp3lxAA;8(FcA%9c?Hp!L$Ln|Y?-An^yi%d;?d{JnG3I?M zC9RZ^r&J0f*iP2z;JEAwVH+c^h5P}lt>3MRx>0doox!;PkJXnFlX@Ec`6 zbeJpt57GXk$~AN%+{I!!WiqLE|C!Y8haE~$+^sqBrT~}E?ytSL*LN=$6zsELi>vG3 zg@t1c)Si92=h5@s3FcVvlzUvJIF6Q{veGp_A5LeT_Ol?FzNrzH+nxV8DJkTLSdctFNaACg;8cD3kwakQaK+Q62dT|nIWBBKkj?TxS)H!k4(QUzSV3Do$PM; zXY7P^e5J4zrrGm0K5^PmZTsDZNpaLB9xLDp8q0+e+$Bh_RGn;wONdTtcf$j57XG3g zT$x#t5Nx#&vUfaOQ)3g}D zWkQP`&3kIuIDfE7fk%qv^gZH!he{p8*sp2-(eqDPsx>2qZw*5R5Nl3 zv0~y?in($kCSMNlehWYD;83QT-R@zvUY9ff7H`nk*Oy9EWSI!=-VD1X>q(;mcT`)_ zP!G^$3Xp!$HW^fA{0d2AIK>S(-aSPAA|d};lEIubC=%2MfKyNz|S zN?>obNzkH`Lm)FQ?IL9VEKV+Oh5eqW7}4NivyOrcHQ=SVqfUAR(Sezcj7_tR3x0>i zZoY~J3H@No8Y~h@UPwtsniI#3Z$dH?ADN01)ku3~we*$TCt$-vi{6tDbKlGBD@<@v zm8$(cp}D>|y5u$^BG8733#9;>sGGFtM+6ltWR}m5n`A0?+?9G$-W1EZTKkBDOMdT2 zRrVT#Ci@5HqKyWT{fre&&u?{q9V-k9eWRf&6$`Kfth7BH~C>E zfEh;kw}s~9LI*H;+9J#@MUh_Q2!e3dtxSN(#fk{j(W)ioFBnI(QGb4%uMnENeUIJt zeU5tzTuRbWtu`gS3S>$QgfnA@`h}9lsVM6+Qs21+_x-yBRHzv9?GxnPQRDNSIGC-} z$WJf);jsy4{4W@doSay*+oxe8-hy=)mFQ9X{C=dGq);vEaQBK<*i%um#IO(v zop}wXdafYSL@aahOy+jl_2Y+TP{pZ3{53|Z1|pdMT}*tL?x0=H8u$YT(i)lACMZ(B zpZNS~giA#9&nL$R_C={iMP(<*@9p6d#tv|hv2e>0b?U4nSMXW>L6<9l zi2kf7VXL|mdCK!OEDINZMJwC%F$81AZ2Pa{9tEXcpdw4kKUf4Ey;%L5cBd~&V@pNf zI?Z1=-s^wZvd3~tTr@7seklJm%v-&di6P$LqEhelUWY|?Ffbt($(RFeL)@9kNLwwS zs?jkKJ?`TsF#P(fI+qXp6Vab<#cyJ9Ss(%-e%NJpXwG;n!XDw!!*ROlLQntV_DHY@ zB*5u^8adr&me&21470K_C7D>!AOH^Dj;5IeqeG|dcZ}NE*qi3-F_=nc1>z>c(NaD0$zbf6BKdwEz1*cE|I!B4cyI5U`~BF{+grxU!vmyx*B(c= zde1N2B(<9_}O0C*gB_eU48z@hIdg7DSj7YGdB< z9P~}&0KdYgx6>Ka^Sy#;$~2^;X?h9cmg{*Pf}1?O{>KHZ;Sh^@@RT%_rDmcYB;o7A zTB#_u+(XbVt|*$C82!%u#&K9Mpez1?r{?YmlyFcTa3-U> zM`)Kh-zM7TN&o%&nn9TaSwriOCYb;7`mn&rI!CEPi8+XXAD-Imbov@>$=ecxKUxeQ zAK^%_4gQ8iPP{U$_=(l{Aq7)7J)k(!l?b(6{EqlT?m!u2K_Ho#U#x%G0IWf@FXqt6 zrQ2~kKi=}86&S5(EI=4?u~gDqsTS)p6ZhoXK~JaB?P-?T&MIv4UT?Tf^mmJ@a`t?@ zOxt!k+<7Q`AK+3z^7xKV8x5+cm{M#d+0R)7j}T6Luv26U14S-7+b(juUl*#q z2x;Nn?M1f(cTA4$nQu`aP{!s4;Z-XX_jd0*90WcD<(8KaM=0F$_n^~&z-?#&)c?oB1=`s2v!}q2e}M!Unr# zA!P>bR>S!rw@g_QHm%@KSs4sY&jVFKLnd;K>rcOSmettdX#<0TY zk2MRBre<7Mw7ug*T}ApOy5!7K433+<%z&JYrp;;nbNPxtg-*Z7cGOrDu7l-oB$!o4dg>@z+MV*f<9hhV&oO)KP!0x5Y|zFq*tIzrFp` z#bS;!=W+!i#?*3YcLi@ETNi%^B}vPSHSr|G>+l}R%lM8G4*&XAjdG8t>wCYC%Ld`% zL||`3(B|Z%Y=*!u(Ii(FI7~;JnTFssJyRzq2JC8vOdhj={H-P%b<)(6-)3E6F9&7# z8e8@LP9=}9r{{3kWAVje4PlrS9PtsRh0Gc?+d?QJ6v|0$um+vg)zz70VFuqmn;^qO zw6!xzSQgL;2xu^1V`7y*T%@0*M1+MQ`|fj`d2sPQ^j7_{1l zIv&sV*=&Z`4)Dfe9Mgun67Ua2O&2i8W!&E2xZr)?d+_t!i%(UJJ zqP=FdbD5FkWcN^n(>an1Qit7)2l2d_a3Y&WLZ>8*n(g)?oDo8N5@)H$=BnDNfw}i% zm{FU<;rjOBt%aDse}^npLcxIB?K7*!K(L4C!rRS0Hp_kr{f3!_`)HPxgO9JPkY_bl zt_O9#zpq~@GX4+Mkm?SiB9FFQs%4a=X%P`P;W)dT;=VB4>FRu^Zgx zH?7xsfMhqoJpekDpWTwfpg zjGFKB$!g&bMoGl>$KY2Fhl>S@Hq)K9_OF@IyYy0N{e#!9OA^z1?fGk6;*@5#SW>)1kOKhBul@04DXwF5^DXW=Ee~%ZgZmH2DX_upalu#r^w&w*e)-Q! z|L}Aj<*{%t*g3~Vi17N-q&Tw1yxcsL{g$8e(OwffJ>E`acGks>nOd2mHULAlN;B-f zayVK*-XYhM(MNL*OU%Fc@cN(vvy|FV^uh++) zH}R~ngy*m!X%_pQ7M3sy18(m5+%CsH>19T6;jtZS>*^g{N6gHv$JqMU#=u!}`lrMn zAosSvY1`OYsaHBGM{QQL5FwqcZcPlFJboOD0(RX)DC{)Z`vocf@!47*-WEq=PQ^p% zw)(zJhELBh+HdQfAMZN^vwv`+WA=nX(y}|~viA?IIUg0cIg;;1{^D_eS=$_J_@R}B z|3tuBCU*zr<$Nqzme|wj_hhSmzdQd_H;=XLx*D8!kVr`Tdw4=Lx!uE+MyDL7{RIlD zK;>;|H<8ulhTDEP@3KceucO}c(iG+};--Nng%;Al;`mYiLYu>G-E1aKP6l|j0ny;S z7zGN~OX_L5FVAHz$wP=LD~5d;q?#BsCdS6c^x6$p>-zqb{=e5f2nXt&CVy4uEb!9j z@w%LISpL>1fJ4#m^ksEj2*_@^;BhnU2o2@Zh@TXX&4blb=iJ;19#+V0s|{Fg2a)^Y z5)vI<{xyzR%IFvvmAz6n+lkkaJsm^>EM$&oAEApSG$*qHiO2 zC(qQzru3R;1xQZT6%zm{dlZ2T2LzS7Rz7&q$GdIe6vhi&qj7{ zUV0(Ct6poh82D@7de9M{yX$?oc02&Z5YCR~qpP5y@p%pMExasvJ>O&c<_|{I3UB-o z6L~Bn;poC&stF2@Evti>+GE`IFGrEKjn5t;DsuevAns9@c4b(u(rVKZQQK_|k2lE3 z&fW@mMlhq5fq&Pj(y5@yP4ssDL8Y_PJ4JNVmWXYj%E~$)MnjFP}+u#-^$R? zariR$g=Uo5>y>N@ZxKOP>50D|Gz!EIyT$J5C*n&$f1uq0$u2QLS&rJa&oM$X_*)6~i!USZ=s_3$LfG+Dt5qGX0xzCsT*viyW@A8o(w@Usl zo^Iq+^q#MKu_Cz}qv6A;3|c9)g~MwFPMbPUdHJ+w3ynmDNKHiI#k`*X%&m-?9S@g8 zelX;Zf8C5DKPsY?b^8m4d`Vbc&Your2_j+`X78N82FpN=6Mo zE|_s{V`5@$4Nay;KWFDqrOR&#-Yk!Z@TL0>WJK{eXWgi?Y&v^bi^=B2_7ndK6b#kl z4|wUepJz~RuB+qr^3(}DrrpkCT&m6(w`29#rO_Ju!aHKrA5VRIJbQnj$LIHU{{lx4 zQHEXR_PJ!CCgyj?5+`o?c#M@nwfBB|e+8*eZDW(`Psc7JcW0xi701X`@YkzCA}quCGfi=JM`%ca(OPP6Z&;RL2{9tG?e`TyLt-FrFKE8E>g$ zv>oNR5=f`XJ#a)Xy?aZi*ZY^Aj{K=!qrA1*FY||mYREEJv7o{0=AS?QOOFyiv4G!Z z*H!6V?WoxJ0wKToE3$`z{bJapPwVa0Ma8abzl$sh%Kg73|E0#dBBG<|hwP`+T-3`* z5ebE4mf!Xf4LaKOgj?@%NF6QJ2ikpB%#|AU+sP8m{UgHJHz9dEF1)H?3*<8xXlPLV zGzz&C>thQ8KOJ&)HFqeb=B11lD)n1pd8R(~Y{C)GRO&qwwb^KhH@c9@hfBYEh=|GH zHxnV0j&D z;?eU)q+|mL5aJ;r0pDG1TYkjNKe_}&1`xkQQ;G_pxV5?OG6n3X+dDk{Q-HBVz~8`+ zf;iTztACJ`3vZTRXQygioOhiI{JZm?$1?sJbIXx$Xd7o)evcw%jKf9Q8dex?1G6_I zg!A@^d*8Y3E8(Wd_hy3smLgCj-|9Tqw;_wk<0u4GD#^aEIuVN1(K%Dfkh=)ZN?6^`g-voHHq5zQ4FJ6Crlwy zgz8YcsG~cM)9A26es~3WfOEwP_7VH1v!5dhDq%(cNuR`M5qK#uu3Fw2O7a(Dk0wWZ zaVF`qtskFXkf2|wxQ@Qjz?JnzsgP#AO+;R|qmK{BEYxi&ECr36_kP_B--)K%f~sE9 z|aP%?fd~ zP{c7R?q#-pALHW)BR;6N+M>qdj$uZ!lxMI?E&dYu!tF5NZXL6!p~BHsf%jp(W`=8^ zf$c4E)ez0-2%=!qy65{X?oW5Djg&GaDJpb=VQRg6eHz^~-^hi)g%MrpqEDaMc3iTN z>rws9)oqGB*y|J^vZOlj$!~w@1MX1e%{k5>@b;v(qXRpc8-@Jw`Eb&2B6B`$Hsxmy zh7s}Nk5KB^@b87FM7d>lQs1sfmC8kDrwFwP`F8sjV~F^7t}=`c7)d-BIvSwaizHY8 z<+Eoz@Lz;vZu@RWg+(1!RQrmDWk4&oQMF>0s1tA2R5$ zI@jk!nZr0NGDY+5LAC&za%$0~oLG8X@%&2ZaI+7t_PDj9QKRdJrF=&Pnbd%4EF*sm zKYd}KS*^Yc8C7p|NQAdHm6xq((#XN(^!R)!8C4xKIxN$2NVRDtzdp9Pu;BVBw`e$E zE|D=5X*C8Va45|Aen&iC&f@jgw0_LuwF~N#BIfg|yPApQ=i_F^7~7V|f-la=K|GGo z?Xve7cBd!w&r=trO3vv>Fda#u))W40;Y5{wR~PVo*RHgjd)t9AoYZU0>fBrFwDJqh zfCWkdi#QOGi1@4Wwce0&J)b#^QueYA8(a<>L%&ii2M#M*K<;Ywkf(yan6~lh*48BF zi>#=rfssuEH$D&MU7f(%%N3W582UVBAFX-u2j`<&Ie+mDW>-kd?ccW^34=%NZJAb8c&Spj z0#)MjdLWzwjvB%bVtM$yj+Q_^;fm!e$Mxj&?@$j{y^Ea4*)R<$DFI$D_ijW!JJ)5P zoMIwnWe6u4j=$vc@1PBooD)&gjq45uPVWkTOpt+GN1sW1H^t)|ZCt_g-g6*(kIS&tX5BU@X@Q)@Fs9R4Ex%mW%BhOe zc6QK*sRp(W9SA0c34SL^2m?oQ+L5W~dy{1N0caZB*6#@QU48;wO}n-R%NLT#kp4N*4bwL2RkJ~&n9u*N{K>~m1PvQfZnX1m_h?R{$Ha#0Q;nU zqrwtGe)lj{qK&R@tz_?n1uQ!s7kYt43kZ(zb&ys4I)4vIxkP<2?&9}zL0(x=fG3I= zzU2IM{!btiFW|wn3j!y~{ppojA{SnP8sg3EST)Hg;P*Kf1Y)P(S7c4ne|Gd5jX@k3 z(U4ZB&=}uK0?1zN_8kHHlcwV0mqqyS7m+OyNeX-3AItFX_YFk-IfazvJb9~1KH`bx7wWv$ogtP zodE()mgm25_koe`-ULQlm9Q-g#4XM888#yE9HnvwjH6~tz$$cEd3$vZB*{g@Ws?~* zf6Den4!!tT+|$bIHIWzRVtK;NY9@1^=Y($R$@Jtu$&crogNFl`Nu7yKp1=PB^f@h* zX`p|VT1u6niGg@K7GFn$P};uJM_Yp~mpkW;me$r*W^@ae3ap9TiPT-Glg%!2>cX#X zuA8p6Gd0>u12%=zCrO~GLF3^h0l0srfWJ2x2|6yW*+e%)I!|~`Y;HUA({^_UG!!KK zckpoDwyH|HjGt&&L^mJd;hJ>K&Gi%6KhgycuQeh?f5q0pgOj}69jYa+!1D_!C^*=b z)s$%D3-WuKfbfT>D8;}3Fb;Gq)r`g_T8J*c-%A{R94*bxNdjrv*pBDJ?ISD#LS%IO zxJjPOzORH;sC(BJQe*9Tm7cH^uqgZ=LyB@{Wyjrs|K_@t$WShbTfE>N=qA=J(;bz! zTaXRHLr0Gua_IsQFWXqu2P2bHybxrqYNr~<~R-0qIY`%O{ik81b>4j-csJD!NRl9&8~ zMfzv=aXBg#=Aqcs2o%W@r}XQAV(Fx@YlwffS@iCQawL3xQ#-LsoTh%$#J20UI%lUi zJu0bS@%w+an`{?p1bGl~wpq{X*Hdb>BWFDyX7gzIyjPz8F3zsr^l$aln%n4Fq)8{{ zt_`nC6cQAyU4v2t7AejJ{c))m6Bjd3WA9MAskXCT~uy*RE}f66Q8H1u{p((Ue*e}U^*Je))P+3kC}2Nu+us}I}IvDNw3 z%FZ*-H2%_`UAs`JY%~!NEg3uaF5q|xqHS^U{*>d(4@0N`aEBgv;vWf}UQTZ+RMyRd z-(1PAJg5|?8(c4@@I6+!2u3kXqb=fSb(yx$Cf_c3J>ekiUPDOLXeeYKK-Qj?k73>F zO>fN#?aIo~G=w%PdVnuT24+ z<|Z`USv22eJN3Pq2%OuDA)CfRvZp%sysn1u$8_fY<57PvI<{h~+m%#*`yy4`N&4K1 z1>m0oGUG-a-lo|>-Fh)Tvhff^!GL&?{wN!rsO(FiUB^{nTC(te%ntn zOTJ|LQZXr@rSR=j zs^4x&Lqi>RyA38*DFw|i!l|3_bk-at9682Nj@irW<|0j6m$!9zuwV6}&Wcy=0NUwb zg@^Xf^5BN=(BWHtBXLY`ySr}h{|-gg)33Jc5BAfEP4`&B(pkg59#v^vjsTP|VE7Bw zXWk!|D^}6dZCmyAKac0wJH4K~#>Z9*p&oQw+yv_O-K+BA;bUSkUCi|=HNhi8T6zz6 z!ZJ=Yb$>;vv{zSF43yDlnRLGI z;N4&Mma|1j?7@jJCHQ5G`bCD5<^KfNxY=m@USA8v_SXi^TTCAwty9Ni3D>f-Tj}_? zTvC$W14H3BcPvk^`>nTX_jk%6RU*#wT{_H}K4u7Wc`YdEglJ`4ilGOFipDp()%Lq2 zrjY(EDxI3wir_C-Fng?uyS5xF;D1b95Ykkr(y>}SMn#O2@*wB6>{0&BXtN~b#jj9za;tF$J;v$ z?USZyas5iT)?BH3x~Jj5C%D1OO(p)22%?}0F}~X&Ad2d(v2EA6+s{KaO&N5Hok>{` zEJy3R@Fo`ss4_ZPvJ$bDZ~kd>sSBPsTdM4f#U4mJUo3Zi+HDw~)>k{4>|SZ|i+1XN z%@wF`6Ek6J`=*SHDGf(yBAAEGu6pb~8I9rmp7flnDNSugV&@MgR2#2Kkl<80Ni#jC+=z7zahIJIXju@zm#c(1?)0Vcm?TX*mO|-$5|Fl6@cvT1zmH&lHNt~zzntzw{p)?YHDJSKxI0B5@wd1wMJIQP! za>MlA-1_Na>yUsmm!pA!p}wHNm6j+@x4de*KlGaGMG!oZU@7uqNh{s`U-02k<1R<} zIl~?`<_Gp6zY+dgG8saJHebsi*iOr06(d?Ut1P39J5=y{5jl3 z143g@16{_pgV_-{haILz6q$PkLG!zJ&~K;eLt`W#HMqAxK_3wrxj1a4Y$R1oWq|~K zx&BoDl5dFNakW$Ay`M0{bYf*)uYT?I{?3LlhbU7b!XN8Z6Y`M>cC1mo-396m>YuWm zX%$`#$$fKa;Q%#5t<}2+^~Eg|6q({M`4+LTU*&+LMA8~zbBUx1iezW|w1rkW(!w&? zizM$HX}Kwal?K`)1!LuK^EP%TD=6W|GV|Ys~$*t-(?+Iy0KshdDI$qf5VSASF`$GAW^JKRpe{(%OGn_Y2)PBxXH`5%3C{E4ZcRb}P?weY zw1LzZNanQa0kydX&@M|(-sH@F)-|9_hONQQMJ6j7t4K%quuCW1P7we5U z+@;u~-XPQWcGVLTiwWi~2nc|B%|wir8l6C$0mLlMXTu7%ajb6}ZI4&l<2_s9@p%e4 ziA+{wjiZHlY^7DNXd}4&edIgeyq-^R({sQ*tQ5a4t8%h%sNnZMpHJ&Ry@wr7#L2oD z^w}J?+s$I@>eheg4j?l1d%Tu6-xT!QyC|(FoR7z5J0zfeso%BycM+sME(Uvw&lW50 zPiHehHt#Gttg(`51L}i3C!dd8qkv-Ah7!-W=h$4P=)8!;#B=HpsPTL1!jwPY$C_RC z>lq-%Z}?{u;15z1$tV zz$YXe7l=?LJ9MYhE9cZ;RPg1zd!6)w(`bC!xk24&3?v(Cs7~pK) zcb&iBoS#^0*lhOf`rdI}pW7z%`NRVUDJyH&>j#8mZ9@{skj=@q5^iEM8K^Z5khgo3XSn(J(t4EVGK!w!4n-9miui@a zqz5KT>{Gc4^RYf$AFoXM5Z+%MciD@7S&7KXdXtCR4tV{v4Fx>TZ*Xuj#9J;;XLQ)E zrsLjPL3>|sD|Zo(Q*H49hO*bTY>kK7-g7ny@}@!QsoHakJ3w`H&CJT^v>MN0wc(xi z<@MdwTF`?FB(?rzbSHLaI)PbxlwzGJm8V$xSUJ)XE7`5P9qDN@+p%KAXm#cFx%b84 z6Eaa>RTOo?^g9L@GQec-JZZx79Pat}G=^Q9j}w8;9d|@brO4thXm3)g-~TB>{nA1;O0T z$bcgX1=~rTC5!GaH(B}^y9(`Q5+B*vnB#^A2pD6@RJd`4cz zvQzf8f(P~y^tmiUm8H+7LpvUBbFE)3AVOt7%#bXFTVg#6J0P88sBPl}Dx|S4wM<)U zb6euciCyAkZ3jbasH|Im93jja2xM@Qs|C2zjrnG}wWn^$r=pHrWVqf3#F5%q*^>Ib z{oRk`q7uD)ZBKzf6ipk6Z_1T9R*5Dp9S62hY92;>+PcYXb53~6Y|$H$TJf*<{wC{$mNHcWd1VUe#aPJ{<)Li zQ22{xS#D*W{(SuB-^y}pcy0dAosczbt*V@)`pk+<@tZ2G=wb{~2_fRIzIWH@%t@-( z+zi@uHcFkD$A-#vlp&cP{IgHxUH=$QO3oUBt zn_*^L?~d@5JVtXfl}6GcCTBKCL!4#2Z5h#NOZnWmA8^P=al|OQf-Lq1qXCcke6@*U zxM@hTX2M8$L_fP@c^W+82@@mfUxuxyR1KL4#c|HWH6g3|=|({X6ZZFVe(e#H5>F&y zLx;EQ>=bm(XTwL!@w|&?Unuyp9#UXFj|kt%4aMBHwu4~J*dUsTx{p{7Hwq~rHQ9(% z!itr$WpAf}LuYqB#*YOY79)c@4kshzOC1(JRA^%%Y)2`QF&Koy^}t@F??gxd-N&)0 z%k}!SumuP|RcI!8GrraRVLVQQ(5Urs?vE?;&I*~V)LQ4p4iNHZwtL?IS*z{6+ZT`} zlhB&W-R(THY1Le#M}Z~Qmpr$9pBO{CRS$&LJo0qi;x;p&l@za(d!)xpPMHZNI~4Lf z@6mBbfRP8$VIbYxr||DCwi+b;w}DIABkFs5N&Ehb`AVI#luyXS0+Y>gV+SRzbPgiI zTe&_;y4VTHF_^FC;?`}^1)+f>LytKw+2C6`vw__2|ep^T@*dBD^oQV(_p>3Vjx|( zpi8&5{st#O#Oh;jHSM*%p?u-31ovFFl2qeGKdLXK309J? zG(`(zN7CXQ|A%6dQtR10c$W~aL?`|lsvz0_FV?AzP>)eRf|OfBLw&O9>qc_qjHusP zz5{Gwmxj2p2*#JZ$1f_E=&hwbQ7IWk&DjJ{otjeOR?;%pE2QP@tLM$!y>IyouxF}EGlf{hK-oQ~>u$Az|C21j;4Z*^}{ znJQ6H`~+EvVeR*!5YoU-MeIT`gGKj~xoY=FL;9Pt5W09$XQ##L*GZ18-Z#l7gNdtE zTn*gQOc0m^qK;1pmt#imyFEZ)qWSr?6@?W&0ekV_GZy=1p~+(v<Q(dKDRufujAd3_k!Q63 z)=Y?*{RcERr^-I5^!E9pyMxjsf)Qu3CrqRn(~mcHDlo8DRYktkC{W4MMgQ+*!IrPb zW*tmogqp4_odi1V zZX}pIM9$^?58WzDe>9jcorw7iW=jAu8exL4?Lv*UC%{xdqaFOHvL9c^YY!eef}x^q zrK76izHa>*bM-~?Q)Zhl$c2sR)hsj)D%mYm+Mlm+KnI#RHlJ|mpJS1BIe>upd;VJ4 zP-_rg&|mhDrQT2+uo*Sqiqmb3l^bLG+bqSlE;SxDI3RZw0TJ$wW24!qNx6QhwMkwW ziI}C3PTkPXzs{U|K!*~!qhe6&MpvI)9}|BT1iY`5_baPDsxtuH|J#zEnN3-h&gE%! zaOQP+z7C+2PXAQ@jU{g29X;26{rsuRY5e|jLpskfUh8i&`Vr2ed<7^?(pG180shRf zu;=&Yuv^11e}`#6xMA4zm=F2FcW6A7QSKyJ-ePkg7QR`}PY7sE`XR6J&EIO~?OogA zX@l!!fcI~>_BE!Qirl685q__W>3~!HGiC1dN|F*;{+8qC*O|uRy=vVqr zChRi5cb$ehNqm(Z;W8GW|%5nBBLgRHwtC`MEEi~G8 z)%)E6kb1`~x65bo@-`qzpb)DjK&>$bV)>F8x@e8Pg`%(LW>WEnt^OXnyEZHd`K+7I z#mj~P{|1*-XS+lg5qyre4d}jZy8a!4sHTas#teVq4v8|71WJ5MOU{w_oN&{m+Qi_XoIUv*NPw z)f%M3cP1ot>GJdDelUnI^OL1O5|e_tp*t2N2`| zI;&z)G6b<&Tbf{Aw^Z z2>?>soNuat_WQ@cN|&R3^eTA!6g z+#91Z)yVND{hKNE>~MZ-f_%Zr-gM3pmw5Taj38}h(9F6>sTr3T&&ciYRhjoYK`Gw{7{ab1BQBx4Sv77 zV7zr*YoH8q&PltW5E`H{bOf|6$vj*(fQ0DgZAJj0egz?ev*;b!=QdB2v$njeYfDz~ zr2v(bR)mm|$9@Ql7TQPtM0SJbX8RLPhS!5Y2qg7B$8V^wSYt`~{d{}sMbE-Clcutl z4DOW5p5}gWV5fjR{GZLmZ0YBA5Di|P{di?8yq-JPXvgBfy<*j!0}v(IY|MFl#~meJ zs&`$?pf+KhsrXd;(X?q6&{nJdZcJ0jUTk3FLsMG>=)d+)oi0fnP2#Vg(8w_o z_k+ipyxp!MJ8hcQT&-gD`5906!*R*$2UtB!OGUlu!&0_uIX`JV#W(kih5DZZJT<1F z7pO5A!nF@a=@D(MLKXU5?y=+I_VZ9jiE927^mO5xE=xuWvTx{oriYj#>Y$N0zf=@5YAp>3?2W3aEQ@5YSiq!l~J|CEzDy>V^+&0{U| zc6{|dGs*r^YQ+dg+79c}QPd(ooBe7XGBtmdi6xzQSEZ!i7>)W$ivzAh+ydxZ)wYNo zR{{cv5YAs|q^?_H;1d&-$;EzWIr8Y!5loL!;vIVNXQu+_yQNMnc>SR}UDZ&UJc1c= zrxl3xMyvCX!-B{V5rWocx0|njMBCV!4zH@S+(bJqAC>vUyhipG6|N7hUy!(5R0+M2 zFjA#B$f~{kzqNp-iu)@2Bd`BlmGJOe3a z;K$vvq5l4#E?Pt=YJl{^Pk2@eO>@(dlaFhm)&kirkOGO)&HBI%(DC%OL^=QZwGD_A zKw?C(&yI0JlEbK7xnsB4vKS+oqRvxeZTeWxMkwXG`3Icv5&UvAl#`;IMSQzC0LtO= zhk>!uYLN*IUU@BHdRnNA-pA znLZ_|9AYLw;1vg=2@oMJarQ2)W(L?Gsu+nWjugU{eM!Cfpjntc<79T4;Ay0n*8>tt zXisndR@G>4rbA@uP-+gH3rXV;eE0{8~ zF`4n1(s8BwKexZ~_ZMDk@2ulLv&f0OXmJoA8**#R!QwrG6J1mxE5CG|kUV0HZq@DV z%gD|*way$xA@xby!*{}CU1@?mjYL}~ff!q>|EaLa0mZ3#oc}-Wfg=@d5Y!sqQQUXF zTJLeQXF`7yMHXGkpE~96^_rXI~trkWW}a?14%j+ z6a)&sfFyby4|ahIqOoE(47kX^9|Ux#t6sOtCW!yL+mv?S$s<(Nn+q5Q9RUxe!%7V2cBk1`*)=;*HJhw2mLaqM=cz`d5e7(4AgCYcjM zxac8lsQkGXN`HoqlxTrMB^Zz4Q?}EETrJ@U5QAHvDrcA%eZ5(W)N(u-v${-SeHhhm zSh48v)Wjy|x9SaDj%tZqaO}B^7M`zJo-6g*>8z9+U!@7hG*LHURi4U_rju?ou35Jh zK%>sQd25jb;&0T*6bh(yeRdYU&g{T#5>$2bc?c}gZ*ri=#v^)vzIJi*UyfyVRJ-F! zzz1Q&HfIW$^wuFQsqZ6Vo;stuI?ips>cg{_c`V|DhuT?`R5~wt ze)>;IQ6n(M@s)f3g4Ylqqxqg>!e8E>Db-^evyMe`jf=A7rboXX=j;w^dNEQNTj8_k zCEW%F4YqodKS!=~ujX40(2kit&Q>3cT`gi~nv72zNzbUIUJ^Sx!`^e zT+Km885tQl+5XLe^q{yK)rzHtSwLf&k7X+-C;xnYZMHdIIBH}$ndW!5jm=xCS-$y5 z&o0{PD<^P6G*{AH9#kMUc=4rtp@$9vSj$}25bo$kbRA610kly1!B;-V<%Atar%k1; z1oicQX~5cjBJM{jo}H;or`KvDu1!orQfIti9VkJ;M4h0RG?Hv=X3yz#$x`}wzD#Ct zsNVaUJ(>3yXeB+v^l}Oc`0Yhf{}zIpw}C|K-_Ohz7+w*QeTUKQhx%hdz6Qr6$OZl5 z`L!Rjs%IE;?^V3ts9}>@yVe#Uq)%^m0fb_?=1RLz)s)($sUGX@YGRZRYrmff&K#*d zn*VL1uFWb32~B#l&C6k>h}3Uhqz50_sD zt*t~wm!r>NBL+=XJEwc$Irpj*;^y;B2D1*L#$5j`T;;<>cR7TrZ+YmQx8?$-wz?pT zl~2ZMb=LY}no`bVkuqy{B_6r?YRmh#%Wd8_d$5PmUj&x;gTKaRCxKR+_t}1WfH4~1 z=`~hOb&coM^~U!-Gh0nY?JE87C(#AzE?vH&X5+12W0gMuHE$B@Cl*5X9S)QAzfXd8 zWg8poI~8;9-w-J?75pSKH?>1DV-ldT>pJ!STCDB%H{oEWm0s1Tv=n#w+?%T{GzeYpcn>5Wn$jk&{R?8*SCL>>Kr z6-~I8k)ke5A!c)+F9iI=Ps5{3Gio3W|Auh4T3LyvXHmozU!SbEwQC&UbWnZZLB7)P zK-76m`=@_aWw`P8Trm>l4F)pjdhS2Wfk}HadkAv@)-R}+r;)0_w!w1D-pe@G3 zWB%n7!r1DolV*1k$1Qx5*`pEBT!Iw8*dLq5EVd$n6m0!U4C{E2IEI1lZI*!}(D2D9 zchGOymeR>lx^o~)v(>I(cvm>^7pY8Am7ouog;#C`5{f)KQ$$6$*(=Y2#@r3);I1SR zut#=BQj=3U-s`m=y&B|5il-pRXpkQ@BolHL3uqI;xTa&_eb9FoplnG1`fVng373MId8r!^&XX9*6_KOC|gh zVKvGRmM}WibO`4z%5iSpudJ@RJepy<+Y;U!u+QX1Z7ey`#HWO8M5gyyP%+QYUb-GFKwzsJb92H_gG``!bRc~&nHc~KB8<%_tY*rP}n zy>GySR}X35`qCw%Mwxy$t@ZClq?wE3>RBo+$f>(s5PjF z`$0#$`umIcf=GjE!5Q~2_hWv*m{)9>o%kgUc=5k1@fBUdlv6WZ!K=^*Zpyg`aJWL@ z?^#T8#gA`ss|EI4ka4X%w*Xk@;3{UiN5c_N77Lb$hXyj_1t&-C-cGHh7(!Wr?M~17fTP&T#Ij&lV#lpA`rE z@+&_ee|Ot^M?KmAD->~5TDDbi7y1-`_jsOTN`umcWT3QHPDg*$>JJ-}i{#>AfoC6F z0R?xz`*T=-KhCI0Xg21ho6FXqS}yqM7b^+x7xm1{lZQ-NAuT+7>N*}3$B2vJSNRm{ z_{58}@P}_LM8p`}dulgHn^)HtIG-Ec7c(CFP5*R9n(87W1OPx%a$%RO^fIgqoS+P2 zuyrD&5m$O311VHaCzUjB(^r9>&%ECIzO4L9x`4@hHb?)G8=; zk=?Ev(->|dlXz>?hB*Ju*8+Fkl>4o{TdH*km|WP)0Vs=N(guA7vcpd5g5&@q4QgDw zx1sT6XfYAIyab%OiZidb0IFVSO!5d|udYhiUTW~mywm37<$D9NSQ=t|cZ3?pR>U$g z#o!t=OOkLUgGLpm=$s;M;q z6#)Lu_>YgHV@6|Q$Nc;m*V8*Lxfg)8v|mW1==9mm*!o=?f-ml^(C!rJ%^QVQ2&8xGA`E=;ctLwOR&6oIqq^JZxZH5KaUeW* zdv$TJ+*~(#cuO0B@BDmzxjp^1Am-%$>QQhop$}){B*fc)Yox5?)cHPf9CIt?8$~fx z;Tw?@8>uUZ+f1P#5KWCjluRy(jkw|8ul2t5Yu(@LTB&QH8+BIl5c4PR$n$MTf}rbM zmn%!0!iC_bm)ni+Ra4p0a=>zf@AoNrS$W}2nBUrPn`z#c(OXBmR4tWp11xm(KYwa< zYVAhSePVb7+5)LkdvZL-yP{eyBF}F*r2%E?*0neMRu&*bF4j7|^THYXaNnUr=0pnk;N=Cq!h9eC#b8>>J;?!i!ry`Ro0CsXV1P&vxh_J{ql%NzlPE*LD8x5Ef9iXw)oo*#JF>m8k4%||`=NzksT4{KzZZjr#c zs1sXmW?H_+^ud~AEzYX17|F!(Rm|!fR#ElvGX}pSlTqhnlZ*Rx0t|8ct%_=fBfByq zvBp8a!yOTVMjn@YDR2}IOsK$>rX_D#y5o%3wZp20b(&a(QftsAd*Xv{Phyk9%!>7HdsCP@yzG)e2ab`Wek|(gee<*dr1;ObduvBWgn*wBoM>Wap2%Hb z?=K&Qv86Oyu9F9lgaP`!B5|~l$&?(R0%b@XrfhCIPyxR4!i#PUAgv?*zx@c3`-8ey z8L+4Q#_PzpmHom5fIP=+Fb{0nW+ob-qJ z-u@FKlqOyo5DSLHM^nb<<*2RdB^YB@F>Jhu?ZDMI0Snnvvt~CDxI@5MlNz9VGw|UY zqQx^WPbsb;QzMWe*9Ktj-CvpFUAaH#cKy$+fVa5a!1M;Am~k|$5OoIQQ9=TIX;gGIhIg!}T=?YpLEqcjLL&;+ z9Cp1A_Vzjwu`nV{UWNfg1}ew-*IQyM$(ccscp(;`f+-pY7d zV;)&oXfR&_wr7?I>}8CFHI<}G;P=madp2wyDJSTdvZYe;GLhXmrZ1;#oU7}D85LML z`?F1y9Mr|Z%7Dxic8?-2*FS+K0Q~v`r4qn=XCNZQ_&Q1TRj70(55z_pV&k7< zU=nIr(E?jx_uvuBpd$lg9-Q%PF=f=X|0sst5|N-G-8Z2zc_9bqIKK=*FlUsZ1qLnr zb#rebzT=VDt7Sm1w7YWqnpZ7=^9)cG7f=hp3_2(@Xi`XmQkS8aBZ~CIW@-pzWMmp< zRO|oF&dyc>*QjvaoPd8hfbhLLr&ULfi`>QBmSz~u%Zyct>{hsYOAU%iB_t+r54K-PW1X^{}^Sfr~|f!H1ww?-D`|d1;jV7Xr)7-lpJ0P!@Rnh z;6)1ukd&I=4BaKHd6M6w0ih|^jLaV{-m$=FlK14)urckrjrH|xvt0c*OMfMd?I<~b zsS6gt;)yRqG8)C?`hEbQCADy~x5)S2awzv`EgZzCha#Z2jo_{uVhDN;r8DvFEGSPA z57X!OQS`yVy1J9MJ_P&IN8l#J01#$WbYT3?Ul}B-Zzo0e5B7*jcCF-g@Z*t%WAK&I zgaszM*tzaLbvA7>(*dV6m*TSlg1e-z@E^Dj2mSm!57C{b1jiwlO|^^uw+yg0%~aS2 z5M>e8Im{sP!oK-W&H%oW%e) zZx?F11NEKf6_U?1xKMRD$FCKMDBC7yY z_VP%_HPc~vYS_@A)7DJ(#F9W>{_^+fu%evFx?!AzYGT{n58!5|ZAK*4{GF`gd>U5- z1!t+Q3IMN%XNFVB6uJXFX4?;?D%{lftV*)Rei6i52Mm0@EKIED?uw)04y-bFI zO5Nk3q|wV05E2p)Ge_rCi~#SKb_zdH1N`*8cOiJjpFC>>epz~Cff)jLf2>p{e9T{p z)}l>Bx_1XA;M_%Wa=synm>M=JGGqt@ra)*ZnY-On06V?`pun7|q-u}9|7xPsO0pz@ z7$G3w;=sm}K)GP?+U~wd%0#3FmT}?&mzTE$b*ut)9A%l~Os)kb&d8 pdiDIAx!9Wuj1WS3S@RKJC2ov#tlfd$0iV25l>I1EE@k}v{{Zg4Zl3@E literal 0 HcmV?d00001 diff --git a/assets/marketing/selectpilot-small-promo.svg b/assets/marketing/selectpilot-small-promo.svg new file mode 100644 index 0000000..4821888 --- /dev/null +++ b/assets/marketing/selectpilot-small-promo.svg @@ -0,0 +1,28 @@ + + + + + + + + + + + + SelectPilot + LOCAL TRANSFORMATION LAYER + + Selected text + to structured output + + + Local-only + + + RESULT + { + "summary": "Action brief", + "actions": ["Ship MVP"], + "risk": "Scope drift" + Privacy proof visible in-panel + diff --git a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md new file mode 100644 index 0000000..7ddce7b --- /dev/null +++ b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md @@ -0,0 +1,67 @@ +# Chrome Web Store Release Checklist + +Last updated: 2026-08-21 + +## Automated Gate + +- [ ] `pnpm lint` +- [ ] `pnpm typecheck` +- [ ] `pnpm test` +- [ ] `pnpm test:privacy` +- [ ] `pnpm build` +- [ ] `pnpm lint:manifest` +- [ ] `pnpm validate:store` +- [ ] `pnpm test:e2e` +- [ ] `pnpm package:store` +- [ ] Package SHA-256 and inventory are preserved with the release evidence. + +`pnpm package:store` must fail while production entitlement verification is absent. Never bypass that failure with a development key or an unsigned entitlement. + +## Product Truth + +- [ ] The listing describes one purpose: selected text to structured local output. +- [ ] Ollama and the local bridge requirements are visible before installation. +- [ ] Trial, paid tiers, and feature boundaries match the exact production configuration. +- [ ] Experimental capabilities are excluded from the core claim. +- [ ] No Team or self-hosted availability is claimed. + +## Privacy And Security + +- [ ] Store disclosures match [`PRIVACY_POLICY.md`](./PRIVACY_POLICY.md). +- [ ] Core content traffic is limited to user-owned localhost services. +- [ ] Any non-content commerce traffic is disclosed separately. +- [ ] Production entitlements are signed and verified fail-closed. +- [ ] No secret, private key, token, log, report, source map, or test artifact exists in the ZIP. +- [ ] Optional retained data can be inspected, exported, and deleted. + +## Assets + +- [ ] Store icon is `128x128`. +- [ ] Three screenshots are `1280x800` and show actual product behavior. +- [ ] Small promo tile is `440x280`. +- [ ] Marquee tile is `1400x560`. +- [ ] Copy and visuals match the submitted extension version. + +## Manual Runtime Acceptance + +- [ ] Fresh Chrome profile installation succeeds. +- [ ] Missing Ollama produces one calm, actionable installation path. +- [ ] Consent-gated model provisioning succeeds on supported hardware. +- [ ] First selected-text extraction reaches a validated result. +- [ ] JSON, Markdown, and plain-text exports open correctly. +- [ ] Failure and recovery paths are verified without silent fallback. +- [ ] Uninstall and retained-data deletion behavior are verified. + +## Dashboard Gates + +- [ ] Homepage URL is public and verified. +- [ ] Support URL is public and verified. +- [ ] Privacy policy URL is public and verified. +- [ ] Listing, Privacy, Distribution, and Test Instructions fields are complete. +- [ ] Reviewer-only setup information contains no repository secret. +- [ ] Saved dashboard fields are reopened and checked for persistence. +- [ ] Submission and publication strategy are explicitly authorized. + +## Completion + +The release is complete only after the exact package is accepted, published, installable from the Store, and reverified through the real first-use path. A local ZIP or green CI run is not publication evidence. diff --git a/docs/CHROME_WEB_STORE_SUBMISSION.md b/docs/CHROME_WEB_STORE_SUBMISSION.md new file mode 100644 index 0000000..f1dd969 --- /dev/null +++ b/docs/CHROME_WEB_STORE_SUBMISSION.md @@ -0,0 +1,109 @@ +# Chrome Web Store Submission Pack + +Last updated: 2026-08-21 + +This document is the source of truth for SelectPilot's store listing. It does not prove that the extension has been submitted or published. + +## Release State + +`BLOCKED`: production entitlement signatures and public policy/support URLs are not yet configured. Do not upload a package until `pnpm package:store` succeeds without overrides. + +## Listing Basics + +- Product name: `SelectPilot` +- Category: `Productivity` +- Primary language: `English` +- Store install: free +- Product access: seven-day trial followed by a paid tier, subject to production checkout and entitlement activation + +## Single Purpose + +SelectPilot turns text you select on a web page into structured, reusable output using a local Ollama runtime. + +## Short Description + +Turn selected text into structured, reusable output with local processing through Ollama. + +## Detailed Description + +SelectPilot helps you turn selected text into useful, consistent output without sending the core transformation flow to a cloud model. + +Select text on a page, open SelectPilot, and choose the result you need: + +- extract people, organizations, dates, decisions, and actions +- create a concise summary +- turn unstructured material into reusable JSON or Markdown +- copy or export the result into another tool + +The core flow runs through a bridge and Ollama on your own computer. SelectPilot shows whether that local runtime is ready before processing begins. + +Important requirements and limits: + +- Ollama and the SelectPilot local bridge are required. +- Initial setup may download a model after explicit user consent. +- Production checkout and signed entitlement issuance must be operational before paid access is advertised as available. +- Experimental audio and vision surfaces are not part of the core selected-text claim. + +## Privacy Tab Mapping + +These answers must be rechecked against the exact submitted SHA and the hosted privacy policy. + +### Data handled for core functionality + +- Website content: selected text and active-page context chosen by the user +- User activity: the user's explicit SelectPilot actions and locally stored feature state +- Authentication information: local entitlement token and signed entitlement metadata + +### Purpose + +- Core functionality only + +### Sale, advertising, and lending + +- Data is not sold. +- Data is not used for personalized advertising. +- Data is not used for creditworthiness or lending. + +### Permission justifications + +- `activeTab`: access the current page only after a user action. +- `tabs`: identify the active content tab for the invoked flow. +- `storage`: retain local settings, entitlement state, runtime state, and user-controlled local history. +- `scripting`: recover the current selection when content-script messaging is unavailable. +- `sidePanel`: provide the primary SelectPilot workspace. +- `http://127.0.0.1/*` and `http://localhost/*`: communicate with user-owned local services only. + +## Store Assets + +- Icon: [`assets/icon128.png`](../assets/icon128.png), `128x128` +- Screenshots: three PNG files in [`assets/marketing`](../assets/marketing), each `1280x800` +- Small promo tile: `selectpilot-small-promo.png`, `440x280` +- Marquee tile: `selectpilot-marquee.png`, `1400x560` + +Run `pnpm validate:store` before uploading any asset. + +## Required Public URLs + +These values are external release gates and must be publicly reachable before submission: + +- Homepage URL: `TBD` +- Support URL: `TBD` +- Privacy policy URL: `TBD` + +The repository copy of the policy is [`docs/PRIVACY_POLICY.md`](./PRIVACY_POLICY.md). A repository file is not a substitute for verifying the final public URL in the Store dashboard. + +## Reviewer Instructions + +1. Install Ollama from its official distribution and start it. +2. Install and start the SelectPilot local bridge. +3. Install the extension package. +4. Follow the in-product setup flow and approve the recommended local model download. +5. Open a web page, select text, and open SelectPilot. +6. Run a structured extraction and export the result. +7. Confirm that the core content flow uses only `127.0.0.1` or `localhost`. + +Reviewer credentials or production entitlement instructions must be supplied through the Store's private reviewer field, never committed to this repository. + +## Submission Rule + +Do not describe SelectPilot as published, purchasable, or fully local until the corresponding behavior has been verified on the exact upload package. Dashboard entry, reviewer approval, and publication remain external actions. diff --git a/docs/PRIVACY_POLICY.md b/docs/PRIVACY_POLICY.md new file mode 100644 index 0000000..ae4746d --- /dev/null +++ b/docs/PRIVACY_POLICY.md @@ -0,0 +1,68 @@ +# SelectPilot Privacy Policy + +Last updated: 2026-08-21 + +SelectPilot turns text selected in a browser into structured output using services running on the user's own computer. + +## Core Processing + +When the user invokes a core text action, SelectPilot may handle: + +- selected text from the active page +- the active page title and URL +- instructions entered in SelectPilot +- generated summaries and structured extraction results + +The extension sends this content to the SelectPilot bridge on `127.0.0.1` or `localhost`. The bridge sends model requests to the user's configured local Ollama runtime. The core selected-text flow has no cloud-model fallback. + +## Local Storage + +SelectPilot may store the following on the user's device: + +- settings and runtime state +- entitlement token and signed entitlement metadata +- model provisioning consent and status +- exported or retained results when the user enables a feature that requires them + +Essential and Plus are designed without retained knowledge history. Pro features that retain local state must make that state visible and provide inspect, export, and delete controls. + +## Commerce And Entitlements + +Payment processing and entitlement issuance are separate from the core text-processing path. Production providers, endpoints, and data handling must be disclosed here before commerce is activated. Reviewer credentials and private tokens are never part of this public policy. + +SelectPilot must not unlock paid features from an unsigned or unverifiable entitlement. The store release remains blocked until production signature verification is configured. + +## Data Sharing + +- Core selected text and generated output are not sold. +- Core selected text and generated output are not used for advertising. +- SelectPilot does not include product telemetry or analytics in the core runtime flow. +- SelectPilot does not send core selected text to an external inference provider by default. + +The user may explicitly export a result to another application or service. That user-directed export is governed by the destination's own privacy terms. + +## Experimental Capabilities + +Audio transcription and vision OCR are experimental and are not included in the core selected-text privacy claim unless their exact runtime path has been separately verified and disclosed. + +## User Control + +Users can: + +- choose when SelectPilot receives selected text +- approve or decline local model downloads +- inspect runtime and privacy state +- inspect, export, or delete supported retained data +- remove local data and uninstall the extension + +## Security + +SelectPilot minimizes data movement and validates local execution boundaries. No software can guarantee absolute security. Security reports should use the public support channel listed in the Chrome Web Store. + +## Public Contact + +`TBD before submission`: publish the same support URL here and in the Chrome Web Store listing. + +## Changes + +Material policy changes must be published no later than the extension version that introduces the corresponding behavior. diff --git a/package.json b/package.json index 7a1853d..da6fcc0 100644 --- a/package.json +++ b/package.json @@ -28,6 +28,8 @@ "build": "pnpm sync:manifest-version && node ./scripts/generate-extraction-presets.mjs && tsc -p tsconfig.json && node ./scripts/build-extension.mjs", "lint": "node ./scripts/generate-extraction-presets.mjs --check && eslint --max-warnings=0 .", "lint:manifest": "node ./scripts/lint-manifest.mjs", + "validate:store": "node ./scripts/validate-store-assets.mjs", + "package:store": "pnpm build && pnpm lint:manifest && node ./scripts/package-chrome-store.mjs", "test:e2e": "playwright test --config tests/e2e/playwright.config.mjs", "test": "tsc -p tsconfig.json && node --test tests/panel/*.test.mjs && python3 -m unittest discover -s tests/server -p 'test_*.py'", "test:privacy": "node --test tests/panel/network-boundary.test.mjs && python3 -m unittest tests/server/test_privacy_proof.py", diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs new file mode 100644 index 0000000..ba203f2 --- /dev/null +++ b/scripts/package-chrome-store.mjs @@ -0,0 +1,111 @@ +#!/usr/bin/env node + +import { createHash } from 'node:crypto'; +import { cp, mkdir, readFile, readdir, rm, stat, utimes, writeFile } from 'node:fs/promises'; +import path from 'node:path'; +import { spawnSync } from 'node:child_process'; +import { fileURLToPath } from 'node:url'; +import { validateStoreAssets } from './validate-store-assets.mjs'; + +const projectRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); +const fixedDate = new Date('2026-01-01T00:00:00Z'); +const runtimeRoots = ['agent', 'api', 'background', 'billing', 'content', 'licensing', 'panel', 'popup', 'pricing', 'shared', 'utils']; +const runtimeAssets = [ + 'assets/icon16.png', + 'assets/icon32.png', + 'assets/icon48.png', + 'assets/icon128.png', + 'assets/icon256.png', + 'assets/icon512.png', +]; +const allowedExtensions = new Set(['.css', '.html', '.js', '.json', '.png', '.svg']); +const forbiddenPath = /(^|\/)(?:\.env|tests?|reports?|logs?|node_modules|test-results|playwright-report)(\/|$)|\.(?:map|log|pem|key)$/i; +const secretPatterns = [ + /-----BEGIN (?:RSA |EC |OPENSSH )?PRIVATE KEY-----/, + /(?:api|secret|private)[_-]?key\s*[:=]\s*["'][^"']{12,}["']/i, +]; + +async function walk(directory, base = directory) { + const entries = await readdir(directory, { withFileTypes: true }); + const files = []; + for (const entry of entries.sort((a, b) => a.name.localeCompare(b.name))) { + const absolute = path.join(directory, entry.name); + if (entry.isDirectory()) files.push(...await walk(absolute, base)); + if (entry.isFile()) files.push(path.relative(base, absolute)); + } + return files; +} + +export async function collectRuntimeFiles(root = projectRoot) { + const files = ['manifest.json', ...runtimeAssets]; + for (const runtimeRoot of runtimeRoots) { + const absoluteRoot = path.join(root, runtimeRoot); + for (const nested of await walk(absoluteRoot)) { + const relative = path.posix.join(runtimeRoot, nested.split(path.sep).join(path.posix.sep)); + if (allowedExtensions.has(path.extname(relative)) && !forbiddenPath.test(relative)) files.push(relative); + } + } + return files.sort(); +} + +export async function assertReleaseSafe(files, root = projectRoot) { + const entitlement = await readFile(path.join(root, 'background/entitlement-service.js'), 'utf8'); + if (/const PUBLIC_KEY_HEX = ['"]{2}/.test(entitlement) || entitlement.includes('accepting as unsigned MVP')) { + throw new Error('Store package blocked: production entitlement signature verification is not configured (SOD-837).'); + } + + for (const relative of files) { + if (forbiddenPath.test(relative)) throw new Error(`Forbidden release path: ${relative}`); + const absolute = path.join(root, relative); + if (!(await stat(absolute)).isFile()) throw new Error(`Missing release file: ${relative}`); + if (!['.png'].includes(path.extname(relative))) { + const content = await readFile(absolute, 'utf8'); + for (const pattern of secretPatterns) { + if (pattern.test(content)) throw new Error(`Potential secret in release file: ${relative}`); + } + } + } +} + +async function sha256(filePath) { + return createHash('sha256').update(await readFile(filePath)).digest('hex'); +} + +export async function packageChromeStore(root = projectRoot) { + await validateStoreAssets(root); + const files = await collectRuntimeFiles(root); + await assertReleaseSafe(files, root); + + const packageJson = JSON.parse(await readFile(path.join(root, 'package.json'), 'utf8')); + const outputRoot = path.join(root, 'dist', 'chrome-web-store'); + const stageRoot = path.join(outputRoot, `selectpilot-${packageJson.version}`); + const zipPath = `${stageRoot}.zip`; + await rm(outputRoot, { recursive: true, force: true }); + await mkdir(stageRoot, { recursive: true }); + + for (const relative of files) { + const destination = path.join(stageRoot, relative); + await mkdir(path.dirname(destination), { recursive: true }); + await cp(path.join(root, relative), destination); + await utimes(destination, fixedDate, fixedDate); + } + + const zip = spawnSync('zip', ['-X', '-q', zipPath, ...files], { cwd: stageRoot, encoding: 'utf8' }); + if (zip.error || zip.status !== 0) throw new Error(`zip failed: ${zip.error?.message || zip.stderr || zip.status}`); + + const report = { + schema_version: 1, + version: packageJson.version, + artifact: path.basename(zipPath), + sha256: await sha256(zipPath), + file_count: files.length, + files, + }; + await writeFile(`${zipPath}.json`, `${JSON.stringify(report, null, 2)}\n`); + return report; +} + +if (process.argv[1] === fileURLToPath(import.meta.url)) { + const report = await packageChromeStore(); + console.log(`${report.artifact} ${report.sha256}`); +} diff --git a/scripts/validate-store-assets.mjs b/scripts/validate-store-assets.mjs new file mode 100644 index 0000000..34354ec --- /dev/null +++ b/scripts/validate-store-assets.mjs @@ -0,0 +1,45 @@ +#!/usr/bin/env node + +import { readFile } from 'node:fs/promises'; +import path from 'node:path'; +import { fileURLToPath } from 'node:url'; + +const projectRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); + +export const REQUIRED_IMAGES = new Map([ + ['assets/icon128.png', [128, 128]], + ['assets/marketing/selectpilot-screenshot-extract.png', [1280, 800]], + ['assets/marketing/selectpilot-screenshot-runtime.png', [1280, 800]], + ['assets/marketing/selectpilot-screenshot-privacy.png', [1280, 800]], + ['assets/marketing/selectpilot-small-promo.png', [440, 280]], + ['assets/marketing/selectpilot-marquee.png', [1400, 560]], +]); + +export function readPngDimensions(buffer) { + const signature = buffer.subarray(0, 8).toString('hex'); + if (signature !== '89504e470d0a1a0a' || buffer.subarray(12, 16).toString('ascii') !== 'IHDR') { + throw new Error('not a PNG with an IHDR header'); + } + return [buffer.readUInt32BE(16), buffer.readUInt32BE(20)]; +} + +export async function validateStoreAssets(root = projectRoot) { + const errors = []; + for (const [relativePath, expected] of REQUIRED_IMAGES) { + try { + const actual = readPngDimensions(await readFile(path.join(root, relativePath))); + if (actual[0] !== expected[0] || actual[1] !== expected[1]) { + errors.push(`${relativePath}: expected ${expected.join('x')}, got ${actual.join('x')}`); + } + } catch (error) { + errors.push(`${relativePath}: ${error.message}`); + } + } + if (errors.length) throw new Error(`Store asset validation failed:\n- ${errors.join('\n- ')}`); + return [...REQUIRED_IMAGES.keys()]; +} + +if (process.argv[1] === fileURLToPath(import.meta.url)) { + await validateStoreAssets(); + console.log(`Validated ${REQUIRED_IMAGES.size} Chrome Web Store images.`); +} diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs new file mode 100644 index 0000000..766957a --- /dev/null +++ b/tests/panel/store-package.test.mjs @@ -0,0 +1,35 @@ +import assert from 'node:assert/strict'; +import test from 'node:test'; +import { readFile } from 'node:fs/promises'; +import path from 'node:path'; +import { fileURLToPath } from 'node:url'; +import { collectRuntimeFiles, assertReleaseSafe } from '../../scripts/package-chrome-store.mjs'; +import { REQUIRED_IMAGES, readPngDimensions, validateStoreAssets } from '../../scripts/validate-store-assets.mjs'; + +const root = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '../..'); + +test('Chrome Web Store image dimensions match the submission contract', async () => { + const files = await validateStoreAssets(root); + assert.deepEqual(files, [...REQUIRED_IMAGES.keys()]); + const icon = await readFile(path.join(root, 'assets/icon128.png')); + assert.deepEqual(readPngDimensions(icon), [128, 128]); +}); + +test('runtime inventory excludes source, tests, reports, and transient files', async () => { + const files = await collectRuntimeFiles(root); + assert.ok(files.includes('manifest.json')); + assert.ok(files.includes('background/background.js')); + assert.ok(files.includes('content/content-script.bundle.js')); + assert.ok(files.includes('assets/icon128.png')); + assert.ok(!files.some((file) => file.startsWith('assets/marketing/'))); + assert.ok(files.every((file) => !file.endsWith('.ts'))); + assert.ok(files.every((file) => !/(^|\/)(tests|reports|node_modules)(\/|$)/.test(file))); +}); + +test('store packaging fails closed while entitlement verification is unsigned', async () => { + const files = await collectRuntimeFiles(root); + await assert.rejects( + assertReleaseSafe(files, root), + /production entitlement signature verification is not configured/ + ); +}); From 671b5c8cf93d2a5b3f704f00ccf1f1105cae55e1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:35:45 +0200 Subject: [PATCH 24/66] fix(governance): map store release evidence Annotate SOD-836 package and validation surfaces for fail-closed specification coverage.\n\nVerification: spec compiler will run on Hetzner exact SHA. Refs: SOD-836 --- scripts/package-chrome-store.mjs | 2 ++ scripts/validate-store-assets.mjs | 3 +++ tests/panel/store-package.test.mjs | 3 +++ 3 files changed, 8 insertions(+) diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs index ba203f2..15826da 100644 --- a/scripts/package-chrome-store.mjs +++ b/scripts/package-chrome-store.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_package-chrome-store_mjs +// spec_ref: "reporting" import { createHash } from 'node:crypto'; import { cp, mkdir, readFile, readdir, rm, stat, utimes, writeFile } from 'node:fs/promises'; diff --git a/scripts/validate-store-assets.mjs b/scripts/validate-store-assets.mjs index 34354ec..05c7034 100644 --- a/scripts/validate-store-assets.mjs +++ b/scripts/validate-store-assets.mjs @@ -1,4 +1,6 @@ #!/usr/bin/env node +// module_name: scripts_validate-store-assets_mjs +// spec_ref: "reporting" import { readFile } from 'node:fs/promises'; import path from 'node:path'; @@ -15,6 +17,7 @@ export const REQUIRED_IMAGES = new Map([ ['assets/marketing/selectpilot-marquee.png', [1400, 560]], ]); +// @spec_ref reporting export function readPngDimensions(buffer) { const signature = buffer.subarray(0, 8).toString('hex'); if (signature !== '89504e470d0a1a0a' || buffer.subarray(12, 16).toString('ascii') !== 'IHDR') { diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs index 766957a..a4012d9 100644 --- a/tests/panel/store-package.test.mjs +++ b/tests/panel/store-package.test.mjs @@ -1,3 +1,6 @@ +// module_name: tests_panel_store-package_test_mjs +// spec_ref: "testing_strategy.integration_tests" + import assert from 'node:assert/strict'; import test from 'node:test'; import { readFile } from 'node:fs/promises'; From c2cb42a29e143463c7c9ef68f202cbc45fbdc24d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 21 Aug 2026 18:37:37 +0200 Subject: [PATCH 25/66] fix(release): make store tooling reproducible Force UTC ZIP metadata and resolve relative CLI entry paths. Add regression coverage proving both commands execute instead of silently no-op.\n\nVerification: node --test tests/panel/store-package.test.mjs Refs: SOD-836 --- scripts/package-chrome-store.mjs | 8 ++++++-- scripts/validate-store-assets.mjs | 2 +- tests/panel/store-package.test.mjs | 16 ++++++++++++++++ 3 files changed, 23 insertions(+), 3 deletions(-) diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs index 15826da..5b549ed 100644 --- a/scripts/package-chrome-store.mjs +++ b/scripts/package-chrome-store.mjs @@ -92,7 +92,11 @@ export async function packageChromeStore(root = projectRoot) { await utimes(destination, fixedDate, fixedDate); } - const zip = spawnSync('zip', ['-X', '-q', zipPath, ...files], { cwd: stageRoot, encoding: 'utf8' }); + const zip = spawnSync('zip', ['-X', '-q', zipPath, ...files], { + cwd: stageRoot, + encoding: 'utf8', + env: { ...process.env, TZ: 'UTC' }, + }); if (zip.error || zip.status !== 0) throw new Error(`zip failed: ${zip.error?.message || zip.stderr || zip.status}`); const report = { @@ -107,7 +111,7 @@ export async function packageChromeStore(root = projectRoot) { return report; } -if (process.argv[1] === fileURLToPath(import.meta.url)) { +if (process.argv[1] && path.resolve(process.argv[1]) === fileURLToPath(import.meta.url)) { const report = await packageChromeStore(); console.log(`${report.artifact} ${report.sha256}`); } diff --git a/scripts/validate-store-assets.mjs b/scripts/validate-store-assets.mjs index 05c7034..321acf5 100644 --- a/scripts/validate-store-assets.mjs +++ b/scripts/validate-store-assets.mjs @@ -42,7 +42,7 @@ export async function validateStoreAssets(root = projectRoot) { return [...REQUIRED_IMAGES.keys()]; } -if (process.argv[1] === fileURLToPath(import.meta.url)) { +if (process.argv[1] && path.resolve(process.argv[1]) === fileURLToPath(import.meta.url)) { await validateStoreAssets(); console.log(`Validated ${REQUIRED_IMAGES.size} Chrome Web Store images.`); } diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs index a4012d9..b55a213 100644 --- a/tests/panel/store-package.test.mjs +++ b/tests/panel/store-package.test.mjs @@ -2,6 +2,7 @@ // spec_ref: "testing_strategy.integration_tests" import assert from 'node:assert/strict'; +import { execFileSync, spawnSync } from 'node:child_process'; import test from 'node:test'; import { readFile } from 'node:fs/promises'; import path from 'node:path'; @@ -36,3 +37,18 @@ test('store packaging fails closed while entitlement verification is unsigned', /production entitlement signature verification is not configured/ ); }); + +test('relative CLI paths execute validation and fail-closed packaging', () => { + const validation = execFileSync(process.execPath, ['./scripts/validate-store-assets.mjs'], { + cwd: root, + encoding: 'utf8', + }); + assert.match(validation, /Validated 6 Chrome Web Store images/); + + const packaging = spawnSync(process.execPath, ['./scripts/package-chrome-store.mjs'], { + cwd: root, + encoding: 'utf8', + }); + assert.equal(packaging.status, 1); + assert.match(packaging.stderr, /production entitlement signature verification is not configured/); +}); From 52bdd27e4f0d3e78d12bbad7a17e1537c4358e5a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 00:52:36 +0200 Subject: [PATCH 26/66] fix(entitlements): require Ed25519 release signatures Reject unsigned, expired, token-mismatched, unknown-key, and tampered entitlements. Keep private signer material outside Git and store packaging blocked until a production public key ring is injected. Verified: pnpm lint, pnpm typecheck, pnpm test, pnpm build, pnpm lint:manifest, pnpm validate:store, negative pnpm package:store, pnpm test:privacy, pnpm test:e2e, pnpm bench:frontier, pnpm runtime:compile-policy, pnpm runtime:validate-policy, node scripts/stress-runner.mjs --- background/entitlement-service.ts | 107 +++++++++++------- docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md | 7 ++ scripts/build-extension.mjs | 26 +++++ scripts/package-chrome-store.mjs | 2 +- server/entitlement_signer.py | 69 +++++++++++ server/monero_payment_server.py | 21 ++-- server/nano_server.py | 14 ++- tests/e2e/extension-first-run-flow.spec.mjs | 46 ++++++-- tests/panel/entitlement-signature.test.mjs | 33 ++++++ tests/server/test_entitlement_signer.py | 58 ++++++++++ tests/server/test_entitlement_verification.py | 12 +- 11 files changed, 328 insertions(+), 67 deletions(-) create mode 100644 server/entitlement_signer.py create mode 100644 tests/panel/entitlement-signature.test.mjs create mode 100644 tests/server/test_entitlement_signer.py diff --git a/background/entitlement-service.ts b/background/entitlement-service.ts index 1a3de4a..0fcfe66 100644 --- a/background/entitlement-service.ts +++ b/background/entitlement-service.ts @@ -41,7 +41,10 @@ export type CachedEntitlement = { const OFFLINE_GRACE_MS = 7 * 24 * 60 * 60 * 1000; const REFRESH_INTERVAL_MS = 10 * 60 * 1000; -const PUBLIC_KEY_HEX = ''; +const ENTITLEMENT_PUBLIC_KEYS: Readonly> = { + __SELECTPILOT_ENTITLEMENT_KEY_ID__: '__SELECTPILOT_ENTITLEMENT_PUBLIC_KEY_HEX__', +}; +const SIGNATURE_ALGORITHM = 'Ed25519'; let cachedFeatureMap: Record | null = null; @@ -92,12 +95,18 @@ function canonicalizeEntitlement(payload: EntitlementPayload): string { }); } -async function verifySignature(payload: EntitlementPayload, signature: string): Promise { - if (!PUBLIC_KEY_HEX) return false; +export async function verifyEntitlementSignature( + payload: EntitlementPayload, + signature: string, + kid: string, + publicKeys: Readonly> = ENTITLEMENT_PUBLIC_KEYS, +): Promise { + const publicKeyHex = publicKeys[kid]; + if (!publicKeyHex || !signature || !kid) return false; try { const key = await crypto.subtle.importKey( 'raw', - bytesToArrayBuffer(hexToBytes(PUBLIC_KEY_HEX)), + bytesToArrayBuffer(hexToBytes(publicKeyHex)), { name: 'Ed25519' }, false, ['verify'] @@ -135,7 +144,25 @@ async function isFeatureAllowedByTier(feature: string, tier: EntitlementTier): P function isWithinOfflineGrace(record: LicenseRecord): boolean { const baseline = record.cachedAt || record.issuedAt; - return nowMs() <= (baseline + OFFLINE_GRACE_MS); + return (!record.expiresAt || nowMs() < record.expiresAt) + && nowMs() <= (baseline + OFFLINE_GRACE_MS); +} + +async function isVerifiedCachedEntitlement(record: LicenseRecord, token: string): Promise { + if ( + record.token !== token + || record.alg !== SIGNATURE_ALGORITHM + || !record.signature + || !record.kid + || !isWithinOfflineGrace(record) + ) return false; + return verifyEntitlementSignature({ + token: record.token, + tier: record.tier, + features: record.features, + issuedAt: record.issuedAt, + expiresAt: record.expiresAt ?? null, + }, record.signature, record.kid); } async function readCachedEntitlement(): Promise { @@ -162,44 +189,38 @@ async function writeCachedEntitlement(record: CachedEntitlement): Promise } async function normalizeRemoteResponse(token: string, response: SignedEntitlementResponse): Promise { - if (response.entitlement) { - const entitlement = response.entitlement; - if (entitlement.token !== token) { - warn('entitlement', 'token mismatch in signed response'); - return null; - } - if (response.signature) { - if (!PUBLIC_KEY_HEX) { - warn('entitlement', 'signature returned but PUBLIC_KEY_HEX is not configured; accepting as unsigned MVP'); - } else { - const valid = await verifySignature(entitlement, response.signature); - if (!valid) return null; - } - } - return { - token: entitlement.token, - tier: entitlement.tier, - features: entitlement.features, - issuedAt: entitlement.issuedAt, - expiresAt: entitlement.expiresAt ?? undefined, - cachedAt: nowMs(), - signature: response.signature, - alg: response.alg, - kid: response.kid, - }; + const entitlement = response.entitlement; + if ( + !entitlement + || entitlement.token !== token + || !['essential', 'plus', 'pro'].includes(entitlement.tier) + || !Number.isInteger(entitlement.issuedAt) + || (entitlement.expiresAt != null && !Number.isInteger(entitlement.expiresAt)) + ) { + warn('entitlement', 'missing or token-mismatched signed entitlement'); + return null; } - - if (response.tier && response.issuedAt) { - return { - token, - tier: response.tier, - issuedAt: response.issuedAt, - expiresAt: response.expiresAt, - cachedAt: nowMs(), - }; + if (response.alg !== SIGNATURE_ALGORITHM || !response.signature || !response.kid) { + warn('entitlement', 'unsigned or unsupported entitlement response'); + return null; } - - return null; + if (entitlement.expiresAt != null && entitlement.expiresAt <= nowMs()) { + warn('entitlement', 'expired entitlement response'); + return null; + } + const valid = await verifyEntitlementSignature(entitlement, response.signature, response.kid); + if (!valid) return null; + return { + token: entitlement.token, + tier: entitlement.tier, + features: entitlement.features, + issuedAt: entitlement.issuedAt, + expiresAt: entitlement.expiresAt ?? undefined, + cachedAt: nowMs(), + signature: response.signature, + alg: response.alg, + kid: response.kid, + }; } async function remoteVerify(token: string): Promise { @@ -245,7 +266,7 @@ export async function refreshEntitlement(force = false): Promise REFRESH_INTERVAL_MS; if (!shouldAttemptRemote && cached) { log('entitlement', 'using cached entitlement within offline grace'); @@ -270,7 +291,7 @@ export async function refreshEntitlement(force = false): Promise !/^[A-Za-z0-9._-]{1,64}$/.test(kid) || !/^[0-9a-f]{64}$/i.test(String(key)))) { + throw new Error('Entitlement public key ring must contain valid key IDs and 32-byte Ed25519 public keys'); + } + const target = path.join(projectRoot, 'background/entitlement-service.js'); + const source = await readFile(target, 'utf8'); + const configured = source.replace( + /const ENTITLEMENT_PUBLIC_KEYS = \{[\s\S]*?\n\};/, + `const ENTITLEMENT_PUBLIC_KEYS = ${JSON.stringify(keys)};`, + ); + if (configured === source) throw new Error('Entitlement public key injection marker not found'); + await writeFile(target, configured); +} + +await injectEntitlementPublicKeys(); + await build({ entryPoints: [path.join(projectRoot, 'content/content-script.ts')], outfile: path.join(projectRoot, 'content/content-script.bundle.js'), diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs index 5b549ed..e9efb97 100644 --- a/scripts/package-chrome-store.mjs +++ b/scripts/package-chrome-store.mjs @@ -52,7 +52,7 @@ export async function collectRuntimeFiles(root = projectRoot) { export async function assertReleaseSafe(files, root = projectRoot) { const entitlement = await readFile(path.join(root, 'background/entitlement-service.js'), 'utf8'); - if (/const PUBLIC_KEY_HEX = ['"]{2}/.test(entitlement) || entitlement.includes('accepting as unsigned MVP')) { + if (entitlement.includes('__SELECTPILOT_ENTITLEMENT_PUBLIC_KEY_HEX__') || entitlement.includes('__SELECTPILOT_ENTITLEMENT_KEY_ID__')) { throw new Error('Store package blocked: production entitlement signature verification is not configured (SOD-837).'); } diff --git a/server/entitlement_signer.py b/server/entitlement_signer.py new file mode 100644 index 0000000..47f2186 --- /dev/null +++ b/server/entitlement_signer.py @@ -0,0 +1,69 @@ +"""module_name: entitlement_signer; spec_ref: "validation_layer".""" + +from __future__ import annotations + +import base64 +import json +import os +import subprocess +import tempfile +from dataclasses import dataclass +from pathlib import Path + + +class SigningError(RuntimeError): + """Explicit fail-closed entitlement signing failure.""" + + +def canonical_entitlement(entitlement: dict) -> bytes: + return json.dumps( + { + "token": entitlement["token"], + "tier": entitlement["tier"], + "features": entitlement.get("features") or [], + "issuedAt": entitlement["issuedAt"], + "expiresAt": entitlement.get("expiresAt"), + }, + separators=(",", ":"), + ensure_ascii=False, + ).encode("utf-8") + + +@dataclass(frozen=True) +class EntitlementSigner: + key_file: str + key_id: str + + @classmethod + def from_environment(cls) -> "EntitlementSigner": + return cls( + key_file=os.environ.get("SELECTPILOT_ENTITLEMENT_SIGNING_KEY_FILE", ""), + key_id=os.environ.get("SELECTPILOT_ENTITLEMENT_SIGNING_KEY_ID", ""), + ) + + def sign(self, entitlement: dict) -> dict: + if not self.key_file or not self.key_id: + raise SigningError("entitlement_signer_not_configured") + key_path = Path(self.key_file) + if not key_path.is_file(): + raise SigningError("entitlement_signing_key_unavailable") + try: + with tempfile.NamedTemporaryFile() as payload_file: + payload_file.write(canonical_entitlement(entitlement)) + payload_file.flush() + result = subprocess.run( + ["openssl", "pkeyutl", "-sign", "-rawin", "-inkey", str(key_path), "-in", payload_file.name], + capture_output=True, + check=False, + timeout=5, + ) + except (OSError, subprocess.TimeoutExpired) as exc: + raise SigningError("entitlement_signing_failed") from exc + if result.returncode != 0 or len(result.stdout) != 64: + raise SigningError("entitlement_signing_failed") + return { + "entitlement": entitlement, + "signature": base64.b64encode(result.stdout).decode("ascii"), + "alg": "Ed25519", + "kid": self.key_id, + } diff --git a/server/monero_payment_server.py b/server/monero_payment_server.py index cd42f3b..0d5d854 100644 --- a/server/monero_payment_server.py +++ b/server/monero_payment_server.py @@ -23,6 +23,7 @@ import requests from flask import Flask, jsonify, request +from entitlement_signer import EntitlementSigner, SigningError app = Flask(__name__) @@ -33,6 +34,7 @@ CONFIRMATIONS_REQUIRED = int(os.environ.get("CHROMEAI_MONERO_CONFIRMATIONS", "10")) POLL_INTERVAL_SECONDS = int(os.environ.get("CHROMEAI_MONERO_POLL_SECONDS", "20")) ORDER_EXPIRY_MS = int(os.environ.get("CHROMEAI_MONERO_ORDER_EXPIRY_MS", str(30 * 60 * 1000))) +ENTITLEMENT_SIGNER = EntitlementSigner.from_environment() db_lock = Lock() @@ -236,13 +238,18 @@ def verify_license(): if not record or record.get("revoked"): return jsonify({"error": "invalid"}), 401 - return jsonify( - { - "tier": record["tier"], - "issuedAt": record["issuedAt"], - "expiresAt": record.get("expiresAt"), - } - ) + entitlement = { + "token": token, + "tier": record["tier"], + "features": record.get("features") or [], + "issuedAt": record["issuedAt"], + "expiresAt": record.get("expiresAt"), + } + try: + return jsonify(ENTITLEMENT_SIGNER.sign(entitlement)) + except SigningError as exc: + app.logger.error("Entitlement signing unavailable: %s", exc) + return jsonify({"error": str(exc)}), 503 # ---- ADMIN REVOKE ---- diff --git a/server/nano_server.py b/server/nano_server.py index b491679..45c43c5 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -1308,14 +1308,22 @@ def license_verify(payload: dict) -> dict: except (URLError, TimeoutError, json.JSONDecodeError) as exc: raise ValidationError("entitlement_verifier_unavailable", "Local entitlement verifier is unavailable", status=503) from exc - issued_at = result.get("issuedAt") if isinstance(result, dict) else None - expires_at = result.get("expiresAt") if isinstance(result, dict) else None + entitlement = result.get("entitlement") if isinstance(result, dict) else None + issued_at = entitlement.get("issuedAt") if isinstance(entitlement, dict) else None + expires_at = entitlement.get("expiresAt") if isinstance(entitlement, dict) else None if ( not isinstance(result, dict) - or result.get("tier") not in {"essential", "plus", "pro"} + or not isinstance(entitlement, dict) + or entitlement.get("token") != token + or entitlement.get("tier") not in {"essential", "plus", "pro"} or not isinstance(issued_at, int) or isinstance(issued_at, bool) or (expires_at is not None and (not isinstance(expires_at, int) or isinstance(expires_at, bool))) + or result.get("alg") != "Ed25519" + or not isinstance(result.get("kid"), str) + or not result.get("kid") + or not isinstance(result.get("signature"), str) + or not result.get("signature") ): raise ValidationError("invalid_entitlement_response", "Local entitlement verifier returned an invalid contract", status=503) return result diff --git a/tests/e2e/extension-first-run-flow.spec.mjs b/tests/e2e/extension-first-run-flow.spec.mjs index 9e8778f..98cde3f 100644 --- a/tests/e2e/extension-first-run-flow.spec.mjs +++ b/tests/e2e/extension-first-run-flow.spec.mjs @@ -1,16 +1,34 @@ // module_name: extension_privacy_integration // spec_ref: "privacy_and_debug_policy" import { test, expect, chromium } from '@playwright/test'; +import { cp, mkdir, readFile, writeFile } from 'node:fs/promises'; +import path from 'node:path'; +import { collectRuntimeFiles } from '../../scripts/package-chrome-store.mjs'; test('real extension preserves privacy from selected text through rendered output', async () => { const executablePath = process.env.SELECTPILOT_CHROME_EXECUTABLE || chromium.executablePath(); + const keyId = 'e2e-ephemeral'; + const keyPair = await crypto.subtle.generateKey('Ed25519', true, ['sign', 'verify']); + const publicKeyHex = Buffer.from(await crypto.subtle.exportKey('raw', keyPair.publicKey)).toString('hex'); + const extensionRoot = test.info().outputPath('extension'); + for (const relative of await collectRuntimeFiles(process.cwd())) { + const destination = path.join(extensionRoot, relative); + await mkdir(path.dirname(destination), { recursive: true }); + await cp(path.join(process.cwd(), relative), destination); + } + const entitlementScript = path.join(extensionRoot, 'background/entitlement-service.js'); + const entitlementSource = await readFile(entitlementScript, 'utf8'); + await writeFile(entitlementScript, entitlementSource.replace( + /const ENTITLEMENT_PUBLIC_KEYS = \{[\s\S]*?\n\};/, + `const ENTITLEMENT_PUBLIC_KEYS = ${JSON.stringify({ [keyId]: publicKeyHex })};`, + )); const context = await chromium.launchPersistentContext(test.info().outputPath('extension-user-data'), { executablePath, headless: process.env.SELECTPILOT_HEADED !== '1', args: [ - `--disable-extensions-except=${process.cwd()}`, - `--load-extension=${process.cwd()}`, + `--disable-extensions-except=${extensionRoot}`, + `--load-extension=${extensionRoot}`, '--disable-web-security', '--disable-background-networking', '--disable-component-update', @@ -74,6 +92,18 @@ test('real extension preserves privacy from selected text through rendered outpu if (path === '/runtime-meta/health') { return fulfill({ ok: false, stream_enabled: false, active_streams: 0, event_version: '1' }); } + if (path === '/license/verify') { + const token = request.postDataJSON().token; + const now = Date.now(); + const entitlement = { + token, tier: 'essential', features: ['structured_extraction'], issuedAt: now, + expiresAt: now + 86_400_000, + }; + const signature = Buffer.from(await crypto.subtle.sign( + 'Ed25519', keyPair.privateKey, new TextEncoder().encode(JSON.stringify(entitlement)), + )).toString('base64'); + return fulfill({ entitlement, signature, alg: 'Ed25519', kid: keyId }); + } if (path === '/extract') { extractRequests.push(request.postDataJSON()); return fulfill({ @@ -107,14 +137,10 @@ test('real extension preserves privacy from selected text through rendered outpu expect(lockedResponse.error).toBe('Paid license required for deterministic extraction'); expect(extractRequests).toEqual([]); - await panelPage.evaluate(async () => { - const storage = await import(globalThis.chrome.runtime.getURL('licensing/license-storage.js')); - const now = Date.now(); - await storage.saveLicense({ - token: 'sp_e2e_paid_token', tier: 'essential', issuedAt: now, - expiresAt: now + 86_400_000, cachedAt: now, - }); - }); + const attached = await panelPage.evaluate(() => globalThis.chrome.runtime.sendMessage({ + type: 'license:attach_token', token: 'sp_e2e_paid_token', + })); + expect(attached).toMatchObject({ token: 'sp_e2e_paid_token', tier: 'essential' }); await panelPage.reload(); await expect(panelPage.locator('#btn-first-run-example')).toBeVisible(); diff --git a/tests/panel/entitlement-signature.test.mjs b/tests/panel/entitlement-signature.test.mjs new file mode 100644 index 0000000..5dbc532 --- /dev/null +++ b/tests/panel/entitlement-signature.test.mjs @@ -0,0 +1,33 @@ +// module_name: tests_panel_entitlement_signature_test_mjs +// spec_ref: "testing_strategy.integration_tests" + +import assert from 'node:assert/strict'; +import test from 'node:test'; +import { verifyEntitlementSignature } from '../../background/entitlement-service.js'; + +function toHex(bytes) { + return Buffer.from(bytes).toString('hex'); +} + +test('only a matching Ed25519 key ID and exact payload verify', async () => { + const pair = await crypto.subtle.generateKey('Ed25519', true, ['sign', 'verify']); + const publicKey = toHex(await crypto.subtle.exportKey('raw', pair.publicKey)); + const entitlement = { + token: 'opaque-token', + tier: 'pro', + features: ['image_ocr'], + issuedAt: 1_700_000_000_000, + expiresAt: 4_000_000_000_000, + }; + const canonical = JSON.stringify(entitlement); + const signature = Buffer.from(await crypto.subtle.sign( + 'Ed25519', + pair.privateKey, + new TextEncoder().encode(canonical), + )).toString('base64'); + const keyRing = { 'rotation-1': publicKey }; + + assert.equal(await verifyEntitlementSignature(entitlement, signature, 'rotation-1', keyRing), true); + assert.equal(await verifyEntitlementSignature(entitlement, signature, 'unknown', keyRing), false); + assert.equal(await verifyEntitlementSignature({ ...entitlement, tier: 'plus' }, signature, 'rotation-1', keyRing), false); +}); diff --git a/tests/server/test_entitlement_signer.py b/tests/server/test_entitlement_signer.py new file mode 100644 index 0000000..b6a0767 --- /dev/null +++ b/tests/server/test_entitlement_signer.py @@ -0,0 +1,58 @@ +"""module_name: entitlement_signer_tests; spec_ref: "testing_strategy.integration_tests".""" + +from __future__ import annotations + +import subprocess +import sys +import tempfile +import unittest +from pathlib import Path + +ROOT = Path(__file__).resolve().parents[2] +SERVER_DIR = ROOT / "server" +if str(SERVER_DIR) not in sys.path: + sys.path.insert(0, str(SERVER_DIR)) + +from entitlement_signer import EntitlementSigner, SigningError, canonical_entitlement # noqa: E402 + + +class EntitlementSignerTests(unittest.TestCase): + entitlement = { + "token": "opaque-token", + "tier": "pro", + "features": ["image_ocr"], + "issuedAt": 1_700_000_000_000, + "expiresAt": 1_700_086_400_000, + } + + def test_unconfigured_signer_fails_closed(self) -> None: + with self.assertRaises(SigningError) as ctx: + EntitlementSigner("", "").sign(self.entitlement) + self.assertEqual(str(ctx.exception), "entitlement_signer_not_configured") + + def test_ephemeral_ed25519_identity_signs_exact_canonical_payload(self) -> None: + with tempfile.TemporaryDirectory() as directory: + key_file = Path(directory) / "signer.pem" + public_file = Path(directory) / "signer-public.pem" + subprocess.run(["openssl", "genpkey", "-algorithm", "Ed25519", "-out", key_file], check=True) + subprocess.run(["openssl", "pkey", "-in", key_file, "-pubout", "-out", public_file], check=True) + + signed = EntitlementSigner(str(key_file), "test-rotation-1").sign(self.entitlement) + signature_file = Path(directory) / "signature.bin" + payload_file = Path(directory) / "payload.json" + import base64 + signature_file.write_bytes(base64.b64decode(signed["signature"])) + payload_file.write_bytes(canonical_entitlement(self.entitlement)) + verified = subprocess.run( + ["openssl", "pkeyutl", "-verify", "-rawin", "-pubin", "-inkey", public_file, + "-sigfile", signature_file, "-in", payload_file], + capture_output=True, + check=False, + ) + self.assertEqual(verified.returncode, 0) + self.assertEqual(signed["alg"], "Ed25519") + self.assertEqual(signed["kid"], "test-rotation-1") + + +if __name__ == "__main__": + unittest.main() diff --git a/tests/server/test_entitlement_verification.py b/tests/server/test_entitlement_verification.py index 21bd2fa..d2f6d48 100644 --- a/tests/server/test_entitlement_verification.py +++ b/tests/server/test_entitlement_verification.py @@ -47,10 +47,13 @@ def test_invalid_authority_token_remains_unauthorized(self) -> None: self.assertEqual(ctx.exception.status, 401) def test_verified_local_authority_response_is_returned(self) -> None: - response = _Response(b'{"tier":"pro","issuedAt":1700000000000,"expiresAt":null}') + response = _Response( + b'{"entitlement":{"token":"opaque-token","tier":"pro","features":[],"issuedAt":1700000000000,' + b'"expiresAt":null},"signature":"c2lnbmF0dXJl","alg":"Ed25519","kid":"rotation-1"}' + ) with patch("nano_server.urlopen", return_value=response): result = license_verify({"token": "opaque-token"}) - self.assertEqual(result["tier"], "pro") + self.assertEqual(result["entitlement"]["tier"], "pro") def test_non_loopback_verifier_configuration_is_rejected(self) -> None: with patch.dict(os.environ, {"SELECTPILOT_BILLING_VERIFY_URL": "https://example.com/license/verify"}): @@ -59,7 +62,10 @@ def test_non_loopback_verifier_configuration_is_rejected(self) -> None: self.assertEqual(ctx.exception.code, "invalid_entitlement_verifier") def test_malformed_authority_response_is_rejected(self) -> None: - response = _Response(b'{"tier":"pro","issuedAt":"not-a-timestamp"}') + response = _Response( + b'{"entitlement":{"token":"opaque-token","tier":"pro","issuedAt":"not-a-timestamp"},' + b'"signature":"bad","alg":"Ed25519","kid":"rotation-1"}' + ) with patch("nano_server.urlopen", return_value=response): with self.assertRaises(ValidationError) as ctx: license_verify({"token": "opaque-token"}) From 928a87d857e39b34003db0fac194c1af9247bd3e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 00:56:15 +0200 Subject: [PATCH 27/66] fix(release): remove inactive remote checkout surface Exclude and delete unreachable Paddle code and placeholder product IDs from the extension, retain only the local entitlement authority, and fail closed when admin revocation is unconfigured. Verified: pnpm lint, pnpm typecheck, pnpm test, pnpm build, pnpm lint:manifest, pnpm validate:store, negative pnpm package:store, pnpm test:privacy, pnpm test:e2e, pnpm bench:frontier, pnpm runtime:compile-policy, pnpm runtime:validate-policy, node scripts/stress-runner.mjs --- README.md | 2 +- background/tier-service.ts | 17 -------------- billing/billing-ui.js | 26 --------------------- billing/billing-ui.ts | 26 --------------------- billing/paddle-checkout.js | 37 ------------------------------ billing/paddle-checkout.ts | 37 ------------------------------ pricing/paddle-products.json | 8 ------- scripts/package-chrome-store.mjs | 6 ++++- server/monero_payment_server.py | 4 +++- tests/panel/store-package.test.mjs | 3 +++ 10 files changed, 12 insertions(+), 154 deletions(-) delete mode 100644 billing/billing-ui.js delete mode 100644 billing/billing-ui.ts delete mode 100644 billing/paddle-checkout.js delete mode 100644 billing/paddle-checkout.ts delete mode 100644 pricing/paddle-products.json diff --git a/README.md b/README.md index 78451ea..bcb4cec 100644 --- a/README.md +++ b/README.md @@ -123,7 +123,7 @@ The repository configuration is authoritative for price and entitlement mapping: Essential and Plus do not retain a knowledge history. Pro stateful features must remain local, visible and user-controlled: retained data can be inspected, exported and deleted. A feature is available only when the signed entitlement and `pricing/tier-feature-map.json` both permit it; there is no silent downgrade or fallback. -Prices and Paddle product identifiers live in `pricing/pricing-global.json` and `pricing/paddle-products.json`. Their presence is configuration evidence, not proof that production checkout or entitlement issuance is operational. +Displayed tier prices live in `pricing/pricing-global.json`. Checkout is not embedded in the extension: paid tokens come from the separately operated local entitlement authority, and Store packaging excludes inactive remote-checkout code and product placeholders. ### Team / self-hosted diff --git a/background/tier-service.ts b/background/tier-service.ts index 546bd6a..8d245f9 100644 --- a/background/tier-service.ts +++ b/background/tier-service.ts @@ -7,14 +7,6 @@ import { setEntitlementToken, } from './entitlement-service.js'; -export type PricingConfig = { - trial: { enabled: boolean; duration_days: number; access: 'pro' | 'plus' | 'essential' }; - tiers: Record<'essential' | 'plus' | 'pro', number>; - offline_grace_days: number; - products?: Record; - vendor_id?: number; -}; - export type FeatureMap = Record<'essential' | 'plus' | 'pro', string[]>; async function loadJSON(path: string): Promise { @@ -23,17 +15,8 @@ async function loadJSON(path: string): Promise { return (await res.json()) as T; } -let cachedPricing: PricingConfig | null = null; let cachedFeatures: FeatureMap | null = null; -export async function getPricing(): Promise { - if (cachedPricing) return cachedPricing; - const pricing = await loadJSON>('pricing/pricing-global.json'); - const productsPayload = await loadJSON<{ products: Record; vendor_id?: number }>('pricing/paddle-products.json'); - cachedPricing = { ...pricing, products: productsPayload.products, vendor_id: productsPayload.vendor_id }; - return cachedPricing; -} - export async function getFeatures(): Promise { if (cachedFeatures) return cachedFeatures; cachedFeatures = await loadJSON('pricing/tier-feature-map.json'); diff --git a/billing/billing-ui.js b/billing/billing-ui.js deleted file mode 100644 index fe2dddc..0000000 --- a/billing/billing-ui.js +++ /dev/null @@ -1,26 +0,0 @@ -// module_name: billing_billing-ui_ts -// spec_ref: "validation_layer" -import { openCheckout } from './paddle-checkout.js'; -import { getPricing } from '../background/tier-service.js'; -import { button, setHTML } from '../utils/dom.js'; -export async function mountBillingUI(root) { - const pricing = await getPricing(); - const tiers = Object.entries(pricing.tiers); - const rows = tiers - .map(([tier, price]) => { - const disabled = pricing.products?.[tier] ? '' : ' disabled'; - return `
${tier}
$${price.toFixed(2)}
`; - }) - .join(''); - setHTML(root, `
${rows}
`); - root.querySelectorAll('button[data-tier]').forEach((btn) => { - btn.addEventListener('click', () => { - const tier = btn.dataset.tier; - const product = pricing.products?.[tier]; - if (!product) - return; - button(btn, true); - openCheckout(product, undefined).finally(() => button(btn, false)); - }); - }); -} diff --git a/billing/billing-ui.ts b/billing/billing-ui.ts deleted file mode 100644 index 7b6861e..0000000 --- a/billing/billing-ui.ts +++ /dev/null @@ -1,26 +0,0 @@ -// module_name: billing_billing-ui_ts -// spec_ref: "validation_layer" -import { openCheckout } from './paddle-checkout.js'; -import { getPricing } from '../background/tier-service.js'; -import { button, setHTML } from '../utils/dom.js'; - -export async function mountBillingUI(root: HTMLElement): Promise { - const pricing = await getPricing(); - const tiers = Object.entries(pricing.tiers); - const rows = tiers - .map(([tier, price]) => { - const disabled = pricing.products?.[tier] ? '' : ' disabled'; - return `
${tier}
$${price.toFixed(2)}
`; - }) - .join(''); - setHTML(root, `
${rows}
`); - root.querySelectorAll('button[data-tier]').forEach((btn) => { - btn.addEventListener('click', () => { - const tier = (btn as HTMLButtonElement).dataset.tier!; - const product = pricing.products?.[tier]; - if (!product) return; - button(btn as HTMLButtonElement, true); - openCheckout(product, undefined).finally(() => button(btn as HTMLButtonElement, false)); - }); - }); -} diff --git a/billing/paddle-checkout.js b/billing/paddle-checkout.js deleted file mode 100644 index 6f26fd9..0000000 --- a/billing/paddle-checkout.js +++ /dev/null @@ -1,37 +0,0 @@ -// module_name: billing_paddle-checkout_ts -// spec_ref: "validation_layer" -import { log, warn } from '../utils/logger.js'; -import { getPricing } from '../background/tier-service.js'; -const PADDLE_SCRIPT = 'https://cdn.paddle.com/paddle/paddle.js'; -async function loadPaddle() { - if (typeof window === 'undefined') - return null; - if (window.Paddle) - return window.Paddle; - await new Promise((resolve, reject) => { - const script = document.createElement('script'); - script.src = PADDLE_SCRIPT; - script.async = true; - script.onload = () => resolve(); - script.onerror = reject; - document.head.appendChild(script); - }); - return window.Paddle; -} -export async function openCheckout(productId, email) { - const Paddle = await loadPaddle(); - if (!Paddle) { - warn('billing', 'Paddle not available'); - return; - } - try { - const pricing = await getPricing(); - const vendor = pricing.vendor_id; - Paddle.Setup({ vendor }); - Paddle.Checkout.open({ product: productId, email, frameTarget: 'paddle-checkout', displayModeTheme: 'light' }); - log('billing', 'opened checkout', productId); - } - catch (e) { - warn('billing', 'checkout failed', e); - } -} diff --git a/billing/paddle-checkout.ts b/billing/paddle-checkout.ts deleted file mode 100644 index d8db105..0000000 --- a/billing/paddle-checkout.ts +++ /dev/null @@ -1,37 +0,0 @@ -// module_name: billing_paddle-checkout_ts -// spec_ref: "validation_layer" -import { log, warn } from '../utils/logger.js'; -import { getPricing } from '../background/tier-service.js'; - -const PADDLE_SCRIPT = 'https://cdn.paddle.com/paddle/paddle.js'; - -async function loadPaddle(): Promise { - if (typeof window === 'undefined') return null; - if ((window as any).Paddle) return (window as any).Paddle; - await new Promise((resolve, reject) => { - const script = document.createElement('script'); - script.src = PADDLE_SCRIPT; - script.async = true; - script.onload = () => resolve(); - script.onerror = reject; - document.head.appendChild(script); - }); - return (window as any).Paddle; -} - -export async function openCheckout(productId: string, email?: string): Promise { - const Paddle = await loadPaddle(); - if (!Paddle) { - warn('billing', 'Paddle not available'); - return; - } - try { - const pricing = await getPricing(); - const vendor = pricing.vendor_id; - Paddle.Setup({ vendor }); - Paddle.Checkout.open({ product: productId, email, frameTarget: 'paddle-checkout', displayModeTheme: 'light' }); - log('billing', 'opened checkout', productId); - } catch (e) { - warn('billing', 'checkout failed', e); - } -} diff --git a/pricing/paddle-products.json b/pricing/paddle-products.json deleted file mode 100644 index 6e4ea0a..0000000 --- a/pricing/paddle-products.json +++ /dev/null @@ -1,8 +0,0 @@ -{ - "products": { - "essential": "prod-essential-199", - "plus": "prod-plus-599", - "pro": "prod-pro-1499" - }, - "vendor_id": 982341 -} diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs index e9efb97..4341837 100644 --- a/scripts/package-chrome-store.mjs +++ b/scripts/package-chrome-store.mjs @@ -11,7 +11,7 @@ import { validateStoreAssets } from './validate-store-assets.mjs'; const projectRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); const fixedDate = new Date('2026-01-01T00:00:00Z'); -const runtimeRoots = ['agent', 'api', 'background', 'billing', 'content', 'licensing', 'panel', 'popup', 'pricing', 'shared', 'utils']; +const runtimeRoots = ['agent', 'api', 'background', 'content', 'licensing', 'panel', 'popup', 'shared', 'utils']; const runtimeAssets = [ 'assets/icon16.png', 'assets/icon32.png', @@ -19,6 +19,7 @@ const runtimeAssets = [ 'assets/icon128.png', 'assets/icon256.png', 'assets/icon512.png', + 'pricing/tier-feature-map.json', ]; const allowedExtensions = new Set(['.css', '.html', '.js', '.json', '.png', '.svg']); const forbiddenPath = /(^|\/)(?:\.env|tests?|reports?|logs?|node_modules|test-results|playwright-report)(\/|$)|\.(?:map|log|pem|key)$/i; @@ -55,6 +56,9 @@ export async function assertReleaseSafe(files, root = projectRoot) { if (entitlement.includes('__SELECTPILOT_ENTITLEMENT_PUBLIC_KEY_HEX__') || entitlement.includes('__SELECTPILOT_ENTITLEMENT_KEY_ID__')) { throw new Error('Store package blocked: production entitlement signature verification is not configured (SOD-837).'); } + if (files.some((file) => file.startsWith('billing/') || file === 'pricing/paddle-products.json')) { + throw new Error('Store package blocked: inactive remote checkout code or product placeholders entered runtime inventory.'); + } for (const relative of files) { if (forbiddenPath.test(relative)) throw new Error(`Forbidden release path: ${relative}`); diff --git a/server/monero_payment_server.py b/server/monero_payment_server.py index 0d5d854..d2f6434 100644 --- a/server/monero_payment_server.py +++ b/server/monero_payment_server.py @@ -30,7 +30,7 @@ # ---- CONFIG ---- RPC_URL = os.environ.get("CHROMEAI_MONERO_RPC_URL", "http://127.0.0.1:18083/json_rpc") DB_FILE = Path(os.environ.get("CHROMEAI_MONERO_DB_FILE", "monero-billing-db.json")) -ADMIN_SECRET = os.environ.get("CHROMEAI_ADMIN_SECRET", "CHANGE_ME") +ADMIN_SECRET = os.environ.get("CHROMEAI_ADMIN_SECRET", "") CONFIRMATIONS_REQUIRED = int(os.environ.get("CHROMEAI_MONERO_CONFIRMATIONS", "10")) POLL_INTERVAL_SECONDS = int(os.environ.get("CHROMEAI_MONERO_POLL_SECONDS", "20")) ORDER_EXPIRY_MS = int(os.environ.get("CHROMEAI_MONERO_ORDER_EXPIRY_MS", str(30 * 60 * 1000))) @@ -255,6 +255,8 @@ def verify_license(): # ---- ADMIN REVOKE ---- @app.post("/admin/revoke") def revoke(): + if not ADMIN_SECRET: + return jsonify({"error": "admin_revoke_not_configured"}), 503 if request.headers.get("x-admin-secret") != ADMIN_SECRET: return jsonify({"error": "forbidden"}), 403 diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs index b55a213..bd57949 100644 --- a/tests/panel/store-package.test.mjs +++ b/tests/panel/store-package.test.mjs @@ -25,6 +25,9 @@ test('runtime inventory excludes source, tests, reports, and transient files', a assert.ok(files.includes('background/background.js')); assert.ok(files.includes('content/content-script.bundle.js')); assert.ok(files.includes('assets/icon128.png')); + assert.ok(files.includes('pricing/tier-feature-map.json')); + assert.ok(!files.some((file) => file.startsWith('billing/'))); + assert.ok(!files.includes('pricing/paddle-products.json')); assert.ok(!files.some((file) => file.startsWith('assets/marketing/'))); assert.ok(files.every((file) => !file.endsWith('.ts'))); assert.ok(files.every((file) => !/(^|\/)(tests|reports|node_modules)(\/|$)/.test(file))); From c6da32c1eeac39cf6931c01dd32e7284e718666d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 00:57:59 +0200 Subject: [PATCH 28/66] fix(bridge): restrict browser origins Reject non-extension browser origins, support exact Store-origin pinning, remove wildcard CORS from JSON and SSE responses, and cap request bodies at 2 MiB. Verified: pnpm lint, pnpm typecheck, pnpm test, pnpm test:privacy, pnpm test:e2e, pnpm bench:frontier, pnpm runtime:compile-policy, pnpm runtime:validate-policy, node scripts/stress-runner.mjs --- README.md | 2 ++ server/nano_server.py | 45 ++++++++++++++++++++++-- tests/server/test_validation_pipeline.py | 15 ++++++++ 3 files changed, 59 insertions(+), 3 deletions(-) diff --git a/README.md b/README.md index bcb4cec..4d45ff7 100644 --- a/README.md +++ b/README.md @@ -67,6 +67,8 @@ tail -n 80 ~/Library/Logs/SelectPilot/nano.log curl -sSf http://127.0.0.1:8083/health ``` +Browser requests to the bridge are accepted only from a Chrome extension origin. Release operators should set `SELECTPILOT_EXTENSION_ORIGIN=chrome-extension://` to pin the exact installed identity; ordinary local command-line health checks omit `Origin` and remain available. + --- ## Architecture diff --git a/server/nano_server.py b/server/nano_server.py index 45c43c5..041cc57 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -9,6 +9,7 @@ import hashlib import json import os +import re import socket import threading import time @@ -27,6 +28,8 @@ DEFAULT_PORT = 8083 LOCAL_HOSTS = {"127.0.0.1", "localhost", "::1"} +MAX_REQUEST_BYTES = 2 * 1024 * 1024 +CHROME_EXTENSION_ORIGIN = re.compile(r"^chrome-extension://[a-p]{32}$") ALLOWED_BRIDGE_ENDPOINT_PATHS = [ "/health", @@ -121,6 +124,16 @@ def __init__(self, code: str, message: str, *, status: int = 400, details: dict self.details = details or {} +def allowed_extension_origin(origin: str | None) -> str | None: + candidate = str(origin or "").strip() + if not candidate: + return None + configured = os.environ.get("SELECTPILOT_EXTENSION_ORIGIN", "").strip() + if configured: + return candidate if candidate == configured and CHROME_EXTENSION_ORIGIN.fullmatch(candidate) else None + return candidate if CHROME_EXTENSION_ORIGIN.fullmatch(candidate) else None + + def _expect_dict(value: object, *, field: str = "payload") -> dict: if not isinstance(value, dict): raise ValidationError( @@ -1334,10 +1347,20 @@ class Handler(BaseHTTPRequestHandler): def _set_headers(self): self.send_header("Content-Type", "application/json") - self.send_header("Access-Control-Allow-Origin", "*") + origin = allowed_extension_origin(self.headers.get("Origin")) + if origin: + self.send_header("Access-Control-Allow-Origin", origin) + self.send_header("Vary", "Origin") self.send_header("Access-Control-Allow-Methods", "GET, POST, OPTIONS") self.send_header("Access-Control-Allow-Headers", "Content-Type, Authorization, Last-Event-ID, x-selectpilot-trace-id, x-trace-id") + def _origin_allowed(self) -> bool: + origin = self.headers.get("Origin") + if not origin or allowed_extension_origin(origin): + return True + self._write_error(403, "origin_not_allowed", "Browser origin is not authorized for the local bridge") + return False + def _write_json(self, status: int, payload: dict): self.send_response(status) self._set_headers() @@ -1357,9 +1380,13 @@ def _write_error(self, status: int, code: str, message: str, details: dict | Non self._write_json(status, payload) def do_OPTIONS(self): + if not self._origin_allowed(): + return self._write_json(204, {}) def do_GET(self): + if not self._origin_allowed(): + return parsed = urlparse(self.path) path = parsed.path.rstrip('/') @@ -1399,7 +1426,10 @@ def do_GET(self): self.send_header("Content-Type", "text/event-stream") self.send_header("Cache-Control", "no-cache") self.send_header("Connection", "keep-alive") - self.send_header("Access-Control-Allow-Origin", "*") + origin = allowed_extension_origin(self.headers.get("Origin")) + if origin: + self.send_header("Access-Control-Allow-Origin", origin) + self.send_header("Vary", "Origin") self.end_headers() RUNTIME_META.increment_streams() @@ -1440,7 +1470,16 @@ def do_GET(self): self._write_json(404, {"error": "not_found"}) def do_POST(self): - length = int(self.headers.get('Content-Length', '0')) + if not self._origin_allowed(): + return + try: + length = int(self.headers.get('Content-Length', '0')) + except ValueError: + self._write_error(400, "invalid_content_length", "Content-Length must be an integer") + return + if length < 0 or length > MAX_REQUEST_BYTES: + self._write_error(413, "request_too_large", f"Request body exceeds {MAX_REQUEST_BYTES} bytes") + return body = self.rfile.read(length) if length else b'{}' try: payload = json.loads(body.decode('utf-8')) diff --git a/tests/server/test_validation_pipeline.py b/tests/server/test_validation_pipeline.py index 5bfd668..28496e5 100644 --- a/tests/server/test_validation_pipeline.py +++ b/tests/server/test_validation_pipeline.py @@ -31,10 +31,25 @@ validate_extract_response, validate_summarize_payload, validate_summarize_response, + allowed_extension_origin, ) class ValidationPipelineTests(unittest.TestCase): + def test_cors_accepts_only_chrome_extension_origins(self) -> None: + extension_origin = "chrome-extension://abcdefghijklmnopabcdefghijklmnop" + self.assertEqual(allowed_extension_origin(extension_origin), extension_origin) + self.assertIsNone(allowed_extension_origin("https://malicious.example")) + self.assertIsNone(allowed_extension_origin("null")) + self.assertIsNone(allowed_extension_origin(None)) + + def test_configured_extension_origin_is_exact(self) -> None: + allowed = "chrome-extension://abcdefghijklmnopabcdefghijklmnop" + other = "chrome-extension://ponmlkjihgfedcbaponmlkjihgfedcba" + with patch.dict("os.environ", {"SELECTPILOT_EXTENSION_ORIGIN": allowed}): + self.assertEqual(allowed_extension_origin(allowed), allowed) + self.assertIsNone(allowed_extension_origin(other)) + def test_runtime_policy_does_not_claim_unverified_promotion(self) -> None: policy = { "policy_version": "test-policy", From 2570d53291ec9932ce6dcb9fc829ca783c8c662d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 01:01:09 +0200 Subject: [PATCH 29/66] fix(billing): harden local entitlement authority Pin the isolated billing runtime, constrain wallet RPC to loopback, use 128-bit order IDs, compare admin secrets in constant time, and persist entitlement state with owner-only permissions. Verified: pnpm lint, pnpm typecheck, pnpm test, pnpm test:billing, pnpm build, pnpm lint:manifest, pnpm validate:store, pnpm test:privacy, pnpm test:e2e, pnpm bench:frontier, pnpm runtime:compile-policy, pnpm runtime:validate-policy, node scripts/stress-runner.mjs, pip check --- .github/workflows/ci.yml | 6 +++ docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md | 2 + package.json | 1 + requirements-billing.txt | 2 + server/billing_security.py | 47 +++++++++++++++++++ server/monero_payment_server.py | 11 ++--- tests/server/test_billing_security.py | 53 ++++++++++++++++++++++ 7 files changed, 116 insertions(+), 6 deletions(-) create mode 100644 requirements-billing.txt create mode 100644 server/billing_security.py create mode 100644 tests/server/test_billing_security.py diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index e01de27..488c97d 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -43,6 +43,9 @@ jobs: - name: Install dependencies run: pnpm install --frozen-lockfile + - name: Install local billing runtime dependencies + run: python -m pip install --requirement requirements-billing.txt + - name: Install Playwright Chromium run: pnpm exec playwright install --with-deps chromium @@ -73,6 +76,9 @@ jobs: - name: Test run: pnpm test + - name: Verify local billing runtime + run: pnpm test:billing + - name: Verify privacy boundary run: pnpm test:privacy diff --git a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md index 0d60e59..bcbfad0 100644 --- a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md +++ b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md @@ -24,6 +24,8 @@ Production packaging requires `SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON` as a JS Never place the private key, its contents, or a production signature fixture in Git, CI logs, Linear, test data, or the extension package. Signer provisioning and activation remain an authorized secrets operation. +The separately operated local billing authority installs its pinned runtime with `python3 -m pip install --requirement requirements-billing.txt`. Its wallet RPC is loopback-only, entitlement/order state is written with owner-only permissions, administrative revocation requires a configured secret of at least 32 characters, and generated order IDs contain 128 bits of random material. + ## Product Truth - [ ] The listing describes one purpose: selected text to structured local output. diff --git a/package.json b/package.json index da6fcc0..d261bec 100644 --- a/package.json +++ b/package.json @@ -33,6 +33,7 @@ "test:e2e": "playwright test --config tests/e2e/playwright.config.mjs", "test": "tsc -p tsconfig.json && node --test tests/panel/*.test.mjs && python3 -m unittest discover -s tests/server -p 'test_*.py'", "test:privacy": "node --test tests/panel/network-boundary.test.mjs && python3 -m unittest tests/server/test_privacy_proof.py", + "test:billing": "python3 -m unittest tests/server/test_billing_security.py tests/server/test_entitlement_signer.py && python3 -c \"import sys; sys.path.insert(0, 'server'); import monero_payment_server\"", "typecheck": "tsc --noEmit -p tsconfig.json", "validate:server": "python3 server/nano_server.py --run-dir /tmp/chromeai-run --log-dir /tmp/chromeai-log" }, diff --git a/requirements-billing.txt b/requirements-billing.txt new file mode 100644 index 0000000..4feeaac --- /dev/null +++ b/requirements-billing.txt @@ -0,0 +1,2 @@ +Flask==3.1.3 +requests==2.34.2 diff --git a/server/billing_security.py b/server/billing_security.py new file mode 100644 index 0000000..b19ad24 --- /dev/null +++ b/server/billing_security.py @@ -0,0 +1,47 @@ +"""module_name: billing_security; spec_ref: "validation_layer".""" + +from __future__ import annotations + +import hmac +import json +import os +import secrets +from pathlib import Path +from urllib.parse import urlparse + + +LOCAL_HOSTS = {"127.0.0.1", "localhost", "::1"} + + +def validated_wallet_rpc_url(value: str) -> str: + parsed = urlparse(value) + if parsed.scheme != "http" or parsed.hostname not in LOCAL_HOSTS or parsed.path != "/json_rpc": + raise RuntimeError("wallet_rpc_must_be_explicit_loopback_http") + return value + + +def new_order_id() -> str: + return "SP-" + secrets.token_hex(16) + + +def admin_secret_matches(configured: str, supplied: str | None) -> bool: + if len(configured) < 32 or supplied is None: + return False + return hmac.compare_digest(configured, supplied) + + +def save_private_json(path: Path, payload: dict) -> None: + path.parent.mkdir(parents=True, exist_ok=True) + temporary = path.with_suffix(f"{path.suffix}.tmp") + descriptor = os.open(temporary, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600) + try: + with os.fdopen(descriptor, "w", encoding="utf-8") as handle: + json.dump(payload, handle, indent=2) + handle.write("\n") + handle.flush() + os.fsync(handle.fileno()) + os.replace(temporary, path) + os.chmod(path, 0o600) + finally: + if temporary.exists(): + temporary.unlink() diff --git a/server/monero_payment_server.py b/server/monero_payment_server.py index d2f6434..d9f7751 100644 --- a/server/monero_payment_server.py +++ b/server/monero_payment_server.py @@ -23,12 +23,13 @@ import requests from flask import Flask, jsonify, request +from billing_security import admin_secret_matches, new_order_id, save_private_json, validated_wallet_rpc_url from entitlement_signer import EntitlementSigner, SigningError app = Flask(__name__) # ---- CONFIG ---- -RPC_URL = os.environ.get("CHROMEAI_MONERO_RPC_URL", "http://127.0.0.1:18083/json_rpc") +RPC_URL = validated_wallet_rpc_url(os.environ.get("CHROMEAI_MONERO_RPC_URL", "http://127.0.0.1:18083/json_rpc")) DB_FILE = Path(os.environ.get("CHROMEAI_MONERO_DB_FILE", "monero-billing-db.json")) ADMIN_SECRET = os.environ.get("CHROMEAI_ADMIN_SECRET", "") CONFIRMATIONS_REQUIRED = int(os.environ.get("CHROMEAI_MONERO_CONFIRMATIONS", "10")) @@ -51,9 +52,7 @@ def load_db(): def save_db(db): - tmp = DB_FILE.with_suffix(".tmp") - tmp.write_text(json.dumps(db, indent=2), encoding="utf-8") - tmp.replace(DB_FILE) + save_private_json(DB_FILE, db) # ---- RPC ---- @@ -99,7 +98,7 @@ def create_order(): if tier not in {"essential", "plus", "pro"}: return jsonify({"error": "invalid_tier"}), 400 - order_id = body.get("order_id") or ("SP-" + secrets.token_hex(4)) + order_id = body.get("order_id") or new_order_id() res = rpc( "create_address", @@ -257,7 +256,7 @@ def verify_license(): def revoke(): if not ADMIN_SECRET: return jsonify({"error": "admin_revoke_not_configured"}), 503 - if request.headers.get("x-admin-secret") != ADMIN_SECRET: + if not admin_secret_matches(ADMIN_SECRET, request.headers.get("x-admin-secret")): return jsonify({"error": "forbidden"}), 403 body = request.get_json(silent=True) or {} diff --git a/tests/server/test_billing_security.py b/tests/server/test_billing_security.py new file mode 100644 index 0000000..08d85c5 --- /dev/null +++ b/tests/server/test_billing_security.py @@ -0,0 +1,53 @@ +"""module_name: billing_security_tests; spec_ref: "testing_strategy.integration_tests".""" + +from __future__ import annotations + +import os +import stat +import sys +import tempfile +import unittest +from pathlib import Path + +ROOT = Path(__file__).resolve().parents[2] +SERVER_DIR = ROOT / "server" +if str(SERVER_DIR) not in sys.path: + sys.path.insert(0, str(SERVER_DIR)) + +from billing_security import ( # noqa: E402 + admin_secret_matches, + new_order_id, + save_private_json, + validated_wallet_rpc_url, +) + + +class BillingSecurityTests(unittest.TestCase): + def test_wallet_rpc_is_loopback_only(self) -> None: + self.assertEqual( + validated_wallet_rpc_url("http://127.0.0.1:18083/json_rpc"), + "http://127.0.0.1:18083/json_rpc", + ) + with self.assertRaises(RuntimeError): + validated_wallet_rpc_url("https://wallet.example/json_rpc") + + def test_order_ids_have_128_bits_of_random_material(self) -> None: + order_id = new_order_id() + self.assertRegex(order_id, r"^SP-[0-9a-f]{32}$") + + def test_admin_secret_requires_configured_high_entropy_value(self) -> None: + secret = "a" * 32 + self.assertTrue(admin_secret_matches(secret, secret)) + self.assertFalse(admin_secret_matches("short", "short")) + self.assertFalse(admin_secret_matches(secret, "b" * 32)) + + def test_billing_database_is_written_owner_only(self) -> None: + with tempfile.TemporaryDirectory() as directory: + path = Path(directory) / "billing.json" + save_private_json(path, {"orders": {}, "entitlements": {}}) + mode = stat.S_IMODE(os.stat(path).st_mode) + self.assertEqual(mode, 0o600) + + +if __name__ == "__main__": + unittest.main() From 8e89fdcaf201f7131e71fa08a36ed531204c8413 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 01:01:54 +0200 Subject: [PATCH 30/66] test(bridge): prove origin and size enforcement Exercise the real HTTP handler for malicious-origin rejection, exact extension-origin CORS, and oversized request rejection. Verified: python3 -m unittest tests/server/test_bridge_origin.py, pnpm bench:frontier --- tests/server/test_bridge_origin.py | 66 ++++++++++++++++++++++++++++++ 1 file changed, 66 insertions(+) create mode 100644 tests/server/test_bridge_origin.py diff --git a/tests/server/test_bridge_origin.py b/tests/server/test_bridge_origin.py new file mode 100644 index 0000000..723c260 --- /dev/null +++ b/tests/server/test_bridge_origin.py @@ -0,0 +1,66 @@ +"""module_name: bridge_origin_tests; spec_ref: "testing_strategy.integration_tests".""" + +from __future__ import annotations + +import http.client +import json +import sys +import threading +import unittest +from http.server import ThreadingHTTPServer +from pathlib import Path + +ROOT = Path(__file__).resolve().parents[2] +SERVER_DIR = ROOT / "server" +if str(SERVER_DIR) not in sys.path: + sys.path.insert(0, str(SERVER_DIR)) + +from nano_server import Handler, MAX_REQUEST_BYTES # noqa: E402 + + +class BridgeOriginTests(unittest.TestCase): + @classmethod + def setUpClass(cls) -> None: + cls.server = ThreadingHTTPServer(("127.0.0.1", 0), Handler) + cls.thread = threading.Thread(target=cls.server.serve_forever, daemon=True) + cls.thread.start() + + @classmethod + def tearDownClass(cls) -> None: + cls.server.shutdown() + cls.server.server_close() + cls.thread.join(timeout=2) + + def request(self, method: str, path: str, *, headers: dict[str, str] | None = None): + connection = http.client.HTTPConnection("127.0.0.1", self.server.server_port, timeout=2) + connection.request(method, path, headers=headers or {}) + response = connection.getresponse() + body = response.read() + response_headers = dict(response.getheaders()) + connection.close() + return response.status, response_headers, body + + def test_rejects_web_origin_without_wildcard_cors(self) -> None: + status, headers, body = self.request("OPTIONS", "/extract", headers={"Origin": "https://malicious.example"}) + self.assertEqual(status, 403) + self.assertNotIn("Access-Control-Allow-Origin", headers) + self.assertEqual(json.loads(body)["error"]["code"], "origin_not_allowed") + + def test_echoes_valid_extension_origin(self) -> None: + origin = "chrome-extension://abcdefghijklmnopabcdefghijklmnop" + status, headers, _body = self.request("OPTIONS", "/extract", headers={"Origin": origin}) + self.assertEqual(status, 204) + self.assertEqual(headers.get("Access-Control-Allow-Origin"), origin) + + def test_rejects_oversized_request_before_reading_body(self) -> None: + status, _headers, body = self.request( + "POST", + "/extract", + headers={"Content-Length": str(MAX_REQUEST_BYTES + 1)}, + ) + self.assertEqual(status, 413) + self.assertEqual(json.loads(body)["error"]["code"], "request_too_large") + + +if __name__ == "__main__": + unittest.main() From 271c628059660a4e704c2a585aeb26732a50abd0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 01:05:24 +0200 Subject: [PATCH 31/66] fix(build): eliminate stale entitlement runtime Remove the unused unsigned verifier and HTML injection helper, commit generated JavaScript matching the hardened TypeScript sources, and fail CI on future TypeScript/JavaScript drift. Verified: pnpm lint, pnpm typecheck, pnpm test, pnpm build, git diff --exit-code -- '*.js', pnpm bench:frontier --- .github/workflows/ci.yml | 3 + background/entitlement-service.js | 119 +++++++++++++++++------------- background/tier-service.js | 9 --- licensing/license-verifier.js | 36 --------- licensing/license-verifier.ts | 41 ---------- utils/dom.js | 5 -- utils/dom.ts | 5 -- 7 files changed, 69 insertions(+), 149 deletions(-) delete mode 100644 licensing/license-verifier.js delete mode 100644 licensing/license-verifier.ts diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 488c97d..3ec62c2 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -67,6 +67,9 @@ jobs: - name: Build run: pnpm build + - name: Verify committed JavaScript matches TypeScript sources + run: git diff --exit-code -- '*.js' + - name: Validate Chrome Web Store assets run: pnpm validate:store diff --git a/background/entitlement-service.js b/background/entitlement-service.js index da1c0c7..d258059 100644 --- a/background/entitlement-service.js +++ b/background/entitlement-service.js @@ -5,7 +5,10 @@ import { error, log, warn } from '../utils/logger.js'; import { loadLicense, loadToken, saveLicense } from '../licensing/license-storage.js'; const OFFLINE_GRACE_MS = 7 * 24 * 60 * 60 * 1000; const REFRESH_INTERVAL_MS = 10 * 60 * 1000; -const PUBLIC_KEY_HEX = ''; +const ENTITLEMENT_PUBLIC_KEYS = { + __SELECTPILOT_ENTITLEMENT_KEY_ID__: '__SELECTPILOT_ENTITLEMENT_PUBLIC_KEY_HEX__', +}; +const SIGNATURE_ALGORITHM = 'Ed25519'; let cachedFeatureMap = null; function nowMs() { return Date.now(); @@ -50,11 +53,12 @@ function canonicalizeEntitlement(payload) { expiresAt: payload.expiresAt ?? null, }); } -async function verifySignature(payload, signature) { - if (!PUBLIC_KEY_HEX) +export async function verifyEntitlementSignature(payload, signature, kid, publicKeys = ENTITLEMENT_PUBLIC_KEYS) { + const publicKeyHex = publicKeys[kid]; + if (!publicKeyHex || !signature || !kid) return false; try { - const key = await crypto.subtle.importKey('raw', bytesToArrayBuffer(hexToBytes(PUBLIC_KEY_HEX)), { name: 'Ed25519' }, false, ['verify']); + const key = await crypto.subtle.importKey('raw', bytesToArrayBuffer(hexToBytes(publicKeyHex)), { name: 'Ed25519' }, false, ['verify']); const valid = await crypto.subtle.verify('Ed25519', key, bytesToArrayBuffer(base64ToBytes(signature)), new TextEncoder().encode(canonicalizeEntitlement(payload))); return Boolean(valid); } @@ -83,7 +87,23 @@ async function isFeatureAllowedByTier(feature, tier) { } function isWithinOfflineGrace(record) { const baseline = record.cachedAt || record.issuedAt; - return nowMs() <= (baseline + OFFLINE_GRACE_MS); + return (!record.expiresAt || nowMs() < record.expiresAt) + && nowMs() <= (baseline + OFFLINE_GRACE_MS); +} +async function isVerifiedCachedEntitlement(record, token) { + if (record.token !== token + || record.alg !== SIGNATURE_ALGORITHM + || !record.signature + || !record.kid + || !isWithinOfflineGrace(record)) + return false; + return verifyEntitlementSignature({ + token: record.token, + tier: record.tier, + features: record.features, + issuedAt: record.issuedAt, + expiresAt: record.expiresAt ?? null, + }, record.signature, record.kid); } async function readCachedEntitlement() { const cached = await loadLicense(); @@ -108,44 +128,37 @@ async function writeCachedEntitlement(record) { }); } async function normalizeRemoteResponse(token, response) { - if (response.entitlement) { - const entitlement = response.entitlement; - if (entitlement.token !== token) { - warn('entitlement', 'token mismatch in signed response'); - return null; - } - if (response.signature) { - if (!PUBLIC_KEY_HEX) { - warn('entitlement', 'signature returned but PUBLIC_KEY_HEX is not configured; accepting as unsigned MVP'); - } - else { - const valid = await verifySignature(entitlement, response.signature); - if (!valid) - return null; - } - } - return { - token: entitlement.token, - tier: entitlement.tier, - features: entitlement.features, - issuedAt: entitlement.issuedAt, - expiresAt: entitlement.expiresAt ?? undefined, - cachedAt: nowMs(), - signature: response.signature, - alg: response.alg, - kid: response.kid, - }; + const entitlement = response.entitlement; + if (!entitlement + || entitlement.token !== token + || !['essential', 'plus', 'pro'].includes(entitlement.tier) + || !Number.isInteger(entitlement.issuedAt) + || (entitlement.expiresAt != null && !Number.isInteger(entitlement.expiresAt))) { + warn('entitlement', 'missing or token-mismatched signed entitlement'); + return null; } - if (response.tier && response.issuedAt) { - return { - token, - tier: response.tier, - issuedAt: response.issuedAt, - expiresAt: response.expiresAt, - cachedAt: nowMs(), - }; + if (response.alg !== SIGNATURE_ALGORITHM || !response.signature || !response.kid) { + warn('entitlement', 'unsigned or unsupported entitlement response'); + return null; } - return null; + if (entitlement.expiresAt != null && entitlement.expiresAt <= nowMs()) { + warn('entitlement', 'expired entitlement response'); + return null; + } + const valid = await verifyEntitlementSignature(entitlement, response.signature, response.kid); + if (!valid) + return null; + return { + token: entitlement.token, + tier: entitlement.tier, + features: entitlement.features, + issuedAt: entitlement.issuedAt, + expiresAt: entitlement.expiresAt ?? undefined, + cachedAt: nowMs(), + signature: response.signature, + alg: response.alg, + kid: response.kid, + }; } async function remoteVerify(token) { try { @@ -168,17 +181,17 @@ async function remoteVerify(token) { } } export async function setEntitlementToken(token) { - const existing = await loadLicense(); + const normalizedToken = token.trim(); + if (!normalizedToken) + throw new Error('entitlement_token_required'); + const resetAt = nowMs(); await saveLicense({ - token, - tier: existing?.tier || 'essential', - issuedAt: existing?.issuedAt || nowMs(), - expiresAt: existing?.expiresAt, - features: existing?.features, - cachedAt: existing?.cachedAt, - signature: existing?.signature, - alg: existing?.alg, - kid: existing?.kid, + token: normalizedToken, + tier: 'essential', + issuedAt: resetAt, + expiresAt: resetAt, + features: [], + cachedAt: 0, }); } export async function refreshEntitlement(force = false) { @@ -188,7 +201,7 @@ export async function refreshEntitlement(force = false) { return null; const shouldAttemptRemote = force || !cached - || !isWithinOfflineGrace(cached) + || !(await isVerifiedCachedEntitlement(cached, token)) || nowMs() - (cached.cachedAt || cached.issuedAt) > REFRESH_INTERVAL_MS; if (!shouldAttemptRemote && cached) { log('entitlement', 'using cached entitlement within offline grace'); @@ -210,7 +223,7 @@ export async function refreshEntitlement(force = false) { await writeCachedEntitlement(remote); return normalizeEntitlement(remote); } - if (cached && isWithinOfflineGrace(cached)) { + if (cached && await isVerifiedCachedEntitlement(cached, token)) { log('entitlement', 'remote unavailable; using cached entitlement within grace window'); return cached; } diff --git a/background/tier-service.js b/background/tier-service.js index e3a776b..aa1b660 100644 --- a/background/tier-service.js +++ b/background/tier-service.js @@ -6,16 +6,7 @@ async function loadJSON(path) { const res = await fetch(url); return (await res.json()); } -let cachedPricing = null; let cachedFeatures = null; -export async function getPricing() { - if (cachedPricing) - return cachedPricing; - const pricing = await loadJSON('pricing/pricing-global.json'); - const productsPayload = await loadJSON('pricing/paddle-products.json'); - cachedPricing = { ...pricing, products: productsPayload.products, vendor_id: productsPayload.vendor_id }; - return cachedPricing; -} export async function getFeatures() { if (cachedFeatures) return cachedFeatures; diff --git a/licensing/license-verifier.js b/licensing/license-verifier.js deleted file mode 100644 index fa41da9..0000000 --- a/licensing/license-verifier.js +++ /dev/null @@ -1,36 +0,0 @@ -// module_name: licensing_license-verifier_ts -// spec_ref: "validation_layer" -import { endpoints } from '../api/endpoints.js'; -import { apiRequest } from '../api/request.js'; -import { daysFrom } from '../utils/time.js'; -import { error, log } from '../utils/logger.js'; -import { loadLicense, loadToken, saveLicense } from './license-storage.js'; -const OFFLINE_GRACE_DAYS = 5; -async function remoteVerify(token) { - try { - const res = await apiRequest(endpoints.licenseVerify, { body: { token } }); - return res; - } - catch (e) { - error('license', 'remote verification failed', e); - return null; - } -} -export async function verifyLicense() { - const cached = await loadLicense(); - const token = await loadToken(); - if (!token) { - return null; - } - const offlineOk = cached && (!cached.expiresAt || Date.now() < cached.expiresAt + OFFLINE_GRACE_DAYS * 24 * 60 * 60 * 1000); - if (offlineOk && daysFrom(cached.issuedAt) <= OFFLINE_GRACE_DAYS) { - log('license', 'using cached license within offline grace'); - return cached; - } - const remote = await remoteVerify(token); - if (remote) { - await saveLicense(remote); - return remote; - } - return offlineOk ? cached : null; -} diff --git a/licensing/license-verifier.ts b/licensing/license-verifier.ts deleted file mode 100644 index fd30076..0000000 --- a/licensing/license-verifier.ts +++ /dev/null @@ -1,41 +0,0 @@ -// module_name: licensing_license-verifier_ts -// spec_ref: "validation_layer" -import { endpoints } from '../api/endpoints.js'; -import { apiRequest } from '../api/request.js'; -import { daysFrom } from '../utils/time.js'; -import { error, log } from '../utils/logger.js'; -import { loadLicense, loadToken, saveLicense, LicenseRecord } from './license-storage.js'; - -const OFFLINE_GRACE_DAYS = 5; - -async function remoteVerify(token: string): Promise { - try { - const res = await apiRequest(endpoints.licenseVerify, { body: { token } }); - return res; - } catch (e) { - error('license', 'remote verification failed', e); - return null; - } -} - -export async function verifyLicense(): Promise { - const cached = await loadLicense(); - const token = await loadToken(); - if (!token) { - return null; - } - - const offlineOk = cached && (!cached.expiresAt || Date.now() < cached.expiresAt + OFFLINE_GRACE_DAYS * 24 * 60 * 60 * 1000); - if (offlineOk && daysFrom(cached.issuedAt) <= OFFLINE_GRACE_DAYS) { - log('license', 'using cached license within offline grace'); - return cached; - } - - const remote = await remoteVerify(token); - if (remote) { - await saveLicense(remote); - return remote; - } - - return offlineOk ? cached : null; -} diff --git a/utils/dom.js b/utils/dom.js index 408ae1f..3fc35fe 100644 --- a/utils/dom.js +++ b/utils/dom.js @@ -14,11 +14,6 @@ export function setText(el, text) { el.textContent = text; } // @spec_ref frontend_state_contract -export function setHTML(el, html) { - if (el) - el.innerHTML = html; -} -// @spec_ref frontend_state_contract export function button(el, disabled) { if (el instanceof HTMLButtonElement) { el.disabled = disabled; diff --git a/utils/dom.ts b/utils/dom.ts index 2d6b417..1d03cc7 100644 --- a/utils/dom.ts +++ b/utils/dom.ts @@ -15,11 +15,6 @@ export function setText(el: HTMLElement | null, text: string): void { if (el) el.textContent = text; } -// @spec_ref frontend_state_contract -export function setHTML(el: HTMLElement | null, html: string): void { - if (el) el.innerHTML = html; -} - // @spec_ref frontend_state_contract export function button(el: HTMLElement | null, disabled: boolean): void { if (el instanceof HTMLButtonElement) { From ec2132cc1c4facbb79984a45eebbc936ec61148c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 01:06:55 +0200 Subject: [PATCH 32/66] fix(security): close CodeQL response leaks Return a stable public signer error and remove user-derived CORS response headers while preserving strict extension-origin admission. Verified: pnpm lint, pnpm typecheck, pnpm test, pnpm test:privacy, pnpm test:e2e, pnpm bench:frontier --- server/monero_payment_server.py | 2 +- server/nano_server.py | 18 +++++------------- tests/server/test_bridge_origin.py | 4 ++-- tests/server/test_validation_pipeline.py | 14 +++++++------- 4 files changed, 15 insertions(+), 23 deletions(-) diff --git a/server/monero_payment_server.py b/server/monero_payment_server.py index d9f7751..dab1316 100644 --- a/server/monero_payment_server.py +++ b/server/monero_payment_server.py @@ -248,7 +248,7 @@ def verify_license(): return jsonify(ENTITLEMENT_SIGNER.sign(entitlement)) except SigningError as exc: app.logger.error("Entitlement signing unavailable: %s", exc) - return jsonify({"error": str(exc)}), 503 + return jsonify({"error": "entitlement_signing_unavailable"}), 503 # ---- ADMIN REVOKE ---- diff --git a/server/nano_server.py b/server/nano_server.py index 041cc57..d119e70 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -124,14 +124,14 @@ def __init__(self, code: str, message: str, *, status: int = 400, details: dict self.details = details or {} -def allowed_extension_origin(origin: str | None) -> str | None: +def extension_origin_allowed(origin: str | None) -> bool: candidate = str(origin or "").strip() if not candidate: - return None + return False configured = os.environ.get("SELECTPILOT_EXTENSION_ORIGIN", "").strip() if configured: - return candidate if candidate == configured and CHROME_EXTENSION_ORIGIN.fullmatch(candidate) else None - return candidate if CHROME_EXTENSION_ORIGIN.fullmatch(candidate) else None + return bool(candidate == configured and CHROME_EXTENSION_ORIGIN.fullmatch(configured)) + return bool(CHROME_EXTENSION_ORIGIN.fullmatch(candidate)) def _expect_dict(value: object, *, field: str = "payload") -> dict: @@ -1347,16 +1347,12 @@ class Handler(BaseHTTPRequestHandler): def _set_headers(self): self.send_header("Content-Type", "application/json") - origin = allowed_extension_origin(self.headers.get("Origin")) - if origin: - self.send_header("Access-Control-Allow-Origin", origin) - self.send_header("Vary", "Origin") self.send_header("Access-Control-Allow-Methods", "GET, POST, OPTIONS") self.send_header("Access-Control-Allow-Headers", "Content-Type, Authorization, Last-Event-ID, x-selectpilot-trace-id, x-trace-id") def _origin_allowed(self) -> bool: origin = self.headers.get("Origin") - if not origin or allowed_extension_origin(origin): + if not origin or extension_origin_allowed(origin): return True self._write_error(403, "origin_not_allowed", "Browser origin is not authorized for the local bridge") return False @@ -1426,10 +1422,6 @@ def do_GET(self): self.send_header("Content-Type", "text/event-stream") self.send_header("Cache-Control", "no-cache") self.send_header("Connection", "keep-alive") - origin = allowed_extension_origin(self.headers.get("Origin")) - if origin: - self.send_header("Access-Control-Allow-Origin", origin) - self.send_header("Vary", "Origin") self.end_headers() RUNTIME_META.increment_streams() diff --git a/tests/server/test_bridge_origin.py b/tests/server/test_bridge_origin.py index 723c260..918be33 100644 --- a/tests/server/test_bridge_origin.py +++ b/tests/server/test_bridge_origin.py @@ -46,11 +46,11 @@ def test_rejects_web_origin_without_wildcard_cors(self) -> None: self.assertNotIn("Access-Control-Allow-Origin", headers) self.assertEqual(json.loads(body)["error"]["code"], "origin_not_allowed") - def test_echoes_valid_extension_origin(self) -> None: + def test_accepts_valid_extension_origin_without_dynamic_cors_header(self) -> None: origin = "chrome-extension://abcdefghijklmnopabcdefghijklmnop" status, headers, _body = self.request("OPTIONS", "/extract", headers={"Origin": origin}) self.assertEqual(status, 204) - self.assertEqual(headers.get("Access-Control-Allow-Origin"), origin) + self.assertNotIn("Access-Control-Allow-Origin", headers) def test_rejects_oversized_request_before_reading_body(self) -> None: status, _headers, body = self.request( diff --git a/tests/server/test_validation_pipeline.py b/tests/server/test_validation_pipeline.py index 28496e5..f369205 100644 --- a/tests/server/test_validation_pipeline.py +++ b/tests/server/test_validation_pipeline.py @@ -31,24 +31,24 @@ validate_extract_response, validate_summarize_payload, validate_summarize_response, - allowed_extension_origin, + extension_origin_allowed, ) class ValidationPipelineTests(unittest.TestCase): def test_cors_accepts_only_chrome_extension_origins(self) -> None: extension_origin = "chrome-extension://abcdefghijklmnopabcdefghijklmnop" - self.assertEqual(allowed_extension_origin(extension_origin), extension_origin) - self.assertIsNone(allowed_extension_origin("https://malicious.example")) - self.assertIsNone(allowed_extension_origin("null")) - self.assertIsNone(allowed_extension_origin(None)) + self.assertTrue(extension_origin_allowed(extension_origin)) + self.assertFalse(extension_origin_allowed("https://malicious.example")) + self.assertFalse(extension_origin_allowed("null")) + self.assertFalse(extension_origin_allowed(None)) def test_configured_extension_origin_is_exact(self) -> None: allowed = "chrome-extension://abcdefghijklmnopabcdefghijklmnop" other = "chrome-extension://ponmlkjihgfedcbaponmlkjihgfedcba" with patch.dict("os.environ", {"SELECTPILOT_EXTENSION_ORIGIN": allowed}): - self.assertEqual(allowed_extension_origin(allowed), allowed) - self.assertIsNone(allowed_extension_origin(other)) + self.assertTrue(extension_origin_allowed(allowed)) + self.assertFalse(extension_origin_allowed(other)) def test_runtime_policy_does_not_claim_unverified_promotion(self) -> None: policy = { From 242e4450622d2c0b12912863fd6eac813c3333e1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 23 Aug 2026 01:14:40 +0200 Subject: [PATCH 33/66] fix(runtime): remove fabricated media capabilities Remove unimplemented transcription and vision surfaces from the extension, bridge, API, documentation, and test harness. Delete the unused legacy nginx proxy and add regression coverage that prevents these capabilities from reappearing without a real runtime. Verified: pnpm lint; pnpm typecheck; pnpm test; pnpm test:billing; pnpm test:privacy; pnpm test:e2e; pnpm bench:frontier; pnpm runtime:validate-policy; node scripts/stress-runner.mjs; pnpm audit --prod --audit-level high --- ROADMAP.md | 4 +- SECURITY.md | 6 +-- VALIDATION_STEPS.md | 5 +-- api/endpoints.js | 2 - api/endpoints.ts | 2 - api/nano-client.js | 6 --- api/nano-client.ts | 10 ----- background/background.js | 29 +-------------- background/background.ts | 27 +------------- docs/CHROME_WEB_STORE_SUBMISSION.md | 2 +- docs/PRIVACY_POLICY.md | 2 +- nginx/chromeai.conf | 53 --------------------------- panel/panel.html | 7 ---- panel/panel.js | 24 ------------ panel/panel.ts | 28 -------------- server/nano_server.py | 31 ---------------- tests/e2e/panel-harness.html | 7 ---- tests/panel/capability-truth.test.mjs | 21 +++++++++++ tests/server/test_capability_truth.py | 23 ++++++++++++ 19 files changed, 54 insertions(+), 235 deletions(-) delete mode 100644 nginx/chromeai.conf create mode 100644 tests/panel/capability-truth.test.mjs create mode 100644 tests/server/test_capability_truth.py diff --git a/ROADMAP.md b/ROADMAP.md index 115b13c..9f7e5c6 100644 --- a/ROADMAP.md +++ b/ROADMAP.md @@ -37,8 +37,8 @@ ### 5) Nginx/hosts-topologi bör vara osynlig -- **Status:** ✅ I huvudsak löst i huvudflödet (direkt mot `127.0.0.1:8083`; nginx markerat som legacy). -- **Åtgärd kvar:** behåll detta som policy i dokumentation och scripts. +- **Status:** ✅ Implementerat i huvudflödet (direkt mot `127.0.0.1:8083`). +- **Implementerat:** den oanvända legacy-nginx-konfigurationen är borttagen så att ingen separat proxy- eller dynamisk CORS-yta kan aktiveras av misstag. ### 6) README saknar tydligt 5-minutersflöde högst upp diff --git a/SECURITY.md b/SECURITY.md index 770fec3..84ea5d7 100644 --- a/SECURITY.md +++ b/SECURITY.md @@ -58,8 +58,8 @@ The following is the core security guarantee of SelectPilot. See `ZERO_LEAKAGE.m ## Known limitations -- `Transcribe` and `Vision OCR` are prototype utilities and are **not** covered by the core privacy boundary claim. -- Billing code exists in the repo as a prototype and is not part of the local selected-text flow. +- Audio transcription and vision OCR are not shipped capabilities. They require separate real local-runtime and privacy qualification before any product surface may expose them. +- The separately operated local entitlement authority is not part of the selected-text inference flow. - The local bridge (`127.0.0.1:8083`) is accessible to any process on the machine. This is an inherent trade-off of local inter-process communication. - SelectPilot does not provide a formal security certification - the privacy boundary is implemented in code and should be treated as a clear product constraint. @@ -76,7 +76,7 @@ Out of scope: - Theoretical vulnerabilities with no realistic exploitation path. - Issues in Ollama itself - report those to the [Ollama project](https://github.com/ollama/ollama). -- Vulnerabilities in prototype surfaces (`Transcribe`, `Vision OCR`, billing) that are clearly marked as out of scope above. +- Unshipped experimental source material that is unreachable from the packaged product. ## Disclosure policy diff --git a/VALIDATION_STEPS.md b/VALIDATION_STEPS.md index 1a36c2e..e8c6d68 100644 --- a/VALIDATION_STEPS.md +++ b/VALIDATION_STEPS.md @@ -33,17 +33,14 @@ - Highlight text on a page. - Trigger buttons: `Extract JSON`, `Summarize`, `Rewrite`, `Action brief`, `Ask Ollama`. - Change the extraction preset and verify the JSON pane and export buttons update. -- Use Advanced tools only for experimental Transcribe/Vision flows. - Confirm the runtime strip shows the active model, ignored remote models, and local-only boundary. ## Endpoints direct - Summarize: `curl http://127.0.0.1:8083/summarize -H 'Content-Type: application/json' -d '{"text":"Sample sentence one. Sample sentence two."}'` - Extract: `curl http://127.0.0.1:8083/extract -H 'Content-Type: application/json' -d '{"preset":"action_brief","text":"Ship beta Friday. Update onboarding copy. Validate launch copy before publishing.","title":"Launch prep","url":"https://example.com"}'` - Agent: `curl http://127.0.0.1:8083/agent -H 'Content-Type: application/json' -d '{"prompt":"Rewrite this as a crisp product pitch.","context":{"selection":"A browser tool that rewrites selected text locally.","url":"https://example.com","title":"Example"}}'` -- Transcribe: `curl http://127.0.0.1:8083/transcribe -H 'Content-Type: application/json' -d '{"audioUrl":"file:///tmp/a.mp3"}'` -- Vision: `curl http://127.0.0.1:8083/vision -H 'Content-Type: application/json' -d '{"imageBase64":"abc"}'` - Embed: `curl http://127.0.0.1:8083/embed -H 'Content-Type: application/json' -d '{"text":"embedding text"}'` -- License: `curl http://127.0.0.1:8083/license/verify -H 'Content-Type: application/json' -d '{"token":"pro-123"}'` +- License verification requires a genuinely signed entitlement from the separately operated local authority; fabricated token examples are intentionally unsupported. - Benchmark: `pnpm benchmark:local` ## Monitoring diff --git a/api/endpoints.js b/api/endpoints.js index 7f0539a..136d323 100644 --- a/api/endpoints.js +++ b/api/endpoints.js @@ -12,8 +12,6 @@ export const endpoints = { privacyProof: `${API_BASE}/privacy-proof`, summarize: `${API_BASE}/summarize`, extract: `${API_BASE}/extract`, - transcribe: `${API_BASE}/transcribe`, - vision: `${API_BASE}/vision`, embed: `${API_BASE}/embed`, agent: `${API_BASE}/agent`, licenseVerify: `${API_BASE}/license/verify`, diff --git a/api/endpoints.ts b/api/endpoints.ts index fe68cfb..4e2ed07 100644 --- a/api/endpoints.ts +++ b/api/endpoints.ts @@ -13,8 +13,6 @@ export const endpoints = { privacyProof: `${API_BASE}/privacy-proof`, summarize: `${API_BASE}/summarize`, extract: `${API_BASE}/extract`, - transcribe: `${API_BASE}/transcribe`, - vision: `${API_BASE}/vision`, embed: `${API_BASE}/embed`, agent: `${API_BASE}/agent`, licenseVerify: `${API_BASE}/license/verify`, diff --git a/api/nano-client.js b/api/nano-client.js index 6282558..b67f2f3 100644 --- a/api/nano-client.js +++ b/api/nano-client.js @@ -8,12 +8,6 @@ export async function summarize(payload) { export async function extract(payload) { return apiRequest(endpoints.extract, { body: payload }); } -export async function transcribe(payload) { - return apiRequest(endpoints.transcribe, { body: payload }); -} -export async function vision(payload) { - return apiRequest(endpoints.vision, { body: payload }); -} export async function embed(payload) { return apiRequest(endpoints.embed, { body: payload }); } diff --git a/api/nano-client.ts b/api/nano-client.ts index 9e792d2..10addfc 100644 --- a/api/nano-client.ts +++ b/api/nano-client.ts @@ -5,8 +5,6 @@ import { apiRequest } from './request.js'; export type SummarizePayload = { text: string; url?: string; title?: string; metadata?: Record }; export type ExtractPayload = { text: string; preset?: string; url?: string; title?: string; metadata?: Record }; -export type TranscribePayload = { audioUrl?: string; mediaId?: string; metadata?: Record }; -export type VisionPayload = { imageBase64?: string; videoFrame?: string; url?: string; metadata?: Record }; export type AgentPayload = { prompt: string; context?: Record }; export type EmbedPayload = { text: string }; export type IntentCompilePayload = { @@ -83,14 +81,6 @@ export async function extract(payload: ExtractPayload) { ); } -export async function transcribe(payload: TranscribePayload) { - return apiRequest<{ text: string; confidence: number }>(endpoints.transcribe, { body: payload }); -} - -export async function vision(payload: VisionPayload) { - return apiRequest<{ text: string; tags?: string[] }>(endpoints.vision, { body: payload }); -} - export async function embed(payload: EmbedPayload) { return apiRequest<{ vector: number[] }>(endpoints.embed, { body: payload }); } diff --git a/background/background.js b/background/background.js index a7db2bf..63dba21 100644 --- a/background/background.js +++ b/background/background.js @@ -1,6 +1,6 @@ // module_name: background_background_ts // spec_ref: "execution_layer" -import { compileIntent, extract, summarize, transcribe, vision } from '../api/nano-client.js'; +import { compileIntent, extract, summarize } from '../api/nano-client.js'; import { runPipeline } from '../agent/agent-pipeline.js'; import { log, error } from '../utils/logger.js'; import { requireFeature, getLicenseTier, attachLicenseToken, refreshLicense } from './tier-service.js'; @@ -210,25 +210,6 @@ async function handleFirstRunExtract() { }, }); } -async function handleTranscribe() { - const { allowed } = await requireFeature('audio_transcription'); - if (!allowed) - throw new Error('Feature blocked: upgrade tier for audio transcription'); - const context = await collectContext(); - if (!context.media?.audio) - throw new Error('No audio element detected on page'); - return transcribe({ audioUrl: context.media.audio, metadata: context.metadata }); -} -async function handleVision() { - const { allowed } = await requireFeature('image_ocr'); - if (!allowed) - throw new Error('Feature blocked: upgrade tier for vision OCR'); - const context = await collectContext(); - const image = context.media?.videoFrame || context.media?.image; - if (!image) - throw new Error('No image or video frame available'); - return vision({ imageBase64: image, url: context.url, metadata: context.metadata }); -} async function handleAgent(prompt) { const { allowed } = await requireFeature('basic_local_agent'); if (!allowed) @@ -280,10 +261,6 @@ chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { sendResponse(await handleSummarize()); return; } - if (msg.type === 'panel:transcribe') { - sendResponse(await handleTranscribe()); - return; - } if (msg.type === 'panel:extract') { sendResponse(await handleExtract(msg.preset)); return; @@ -292,10 +269,6 @@ chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { sendResponse(await handleFirstRunExtract()); return; } - if (msg.type === 'panel:vision') { - sendResponse(await handleVision()); - return; - } if (msg.type === 'panel:agent') { sendResponse(await handleAgent(msg.prompt)); return; diff --git a/background/background.ts b/background/background.ts index 068703c..be4d3b7 100644 --- a/background/background.ts +++ b/background/background.ts @@ -1,6 +1,6 @@ // module_name: background_background_ts // spec_ref: "execution_layer" -import { compileIntent, extract, summarize, transcribe, vision } from '../api/nano-client.js'; +import { compileIntent, extract, summarize } from '../api/nano-client.js'; import { runPipeline } from '../agent/agent-pipeline.js'; import { log, error } from '../utils/logger.js'; import { requireFeature, getLicenseTier, attachLicenseToken, refreshLicense } from './tier-service.js'; @@ -239,23 +239,6 @@ async function handleFirstRunExtract(): Promise { }); } -async function handleTranscribe(): Promise { - const { allowed } = await requireFeature('audio_transcription'); - if (!allowed) throw new Error('Feature blocked: upgrade tier for audio transcription'); - const context = await collectContext(); - if (!context.media?.audio) throw new Error('No audio element detected on page'); - return transcribe({ audioUrl: context.media.audio, metadata: context.metadata }); -} - -async function handleVision(): Promise { - const { allowed } = await requireFeature('image_ocr'); - if (!allowed) throw new Error('Feature blocked: upgrade tier for vision OCR'); - const context = await collectContext(); - const image = context.media?.videoFrame || context.media?.image; - if (!image) throw new Error('No image or video frame available'); - return vision({ imageBase64: image, url: context.url, metadata: context.metadata }); -} - async function handleAgent(prompt: string): Promise { const { allowed } = await requireFeature('basic_local_agent'); if (!allowed) throw new Error('Flow tier required for ask/rewrite transforms'); @@ -309,10 +292,6 @@ chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { sendResponse(await handleSummarize()); return; } - if (msg.type === 'panel:transcribe') { - sendResponse(await handleTranscribe()); - return; - } if (msg.type === 'panel:extract') { sendResponse(await handleExtract(msg.preset)); return; @@ -321,10 +300,6 @@ chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { sendResponse(await handleFirstRunExtract()); return; } - if (msg.type === 'panel:vision') { - sendResponse(await handleVision()); - return; - } if (msg.type === 'panel:agent') { sendResponse(await handleAgent(msg.prompt)); return; diff --git a/docs/CHROME_WEB_STORE_SUBMISSION.md b/docs/CHROME_WEB_STORE_SUBMISSION.md index f1dd969..06a96f5 100644 --- a/docs/CHROME_WEB_STORE_SUBMISSION.md +++ b/docs/CHROME_WEB_STORE_SUBMISSION.md @@ -42,7 +42,7 @@ Important requirements and limits: - Ollama and the SelectPilot local bridge are required. - Initial setup may download a model after explicit user consent. - Production checkout and signed entitlement issuance must be operational before paid access is advertised as available. -- Experimental audio and vision surfaces are not part of the core selected-text claim. +- Audio transcription and vision OCR are not shipped capabilities. ## Privacy Tab Mapping diff --git a/docs/PRIVACY_POLICY.md b/docs/PRIVACY_POLICY.md index ae4746d..65dfe53 100644 --- a/docs/PRIVACY_POLICY.md +++ b/docs/PRIVACY_POLICY.md @@ -43,7 +43,7 @@ The user may explicitly export a result to another application or service. That ## Experimental Capabilities -Audio transcription and vision OCR are experimental and are not included in the core selected-text privacy claim unless their exact runtime path has been separately verified and disclosed. +Audio transcription and vision OCR are not shipped capabilities. They require separate real local-runtime and privacy qualification before they can be exposed. ## User Control diff --git a/nginx/chromeai.conf b/nginx/chromeai.conf deleted file mode 100644 index 840472e..0000000 --- a/nginx/chromeai.conf +++ /dev/null @@ -1,53 +0,0 @@ -worker_processes 1; -error_log /usr/local/var/log/chromeai/nginx-error.log; -pid /usr/local/var/run/chromeai/nginx.pid; - -events { worker_connections 1024; } - -http { - include mime.types; - default_type application/json; - sendfile on; - keepalive_timeout 65; - - map $http_origin $cors_origin { - default ""; - ~^chrome-extension://.* $http_origin; - } - - map $request_uri $nano_path { - default $request_uri; - } - - set $nano_port ""; - include /usr/local/var/run/chromeai/port.info; - - if ($nano_port = "") { return 503; } - if ($nano_port !~ ^80(8[0-9]|9[0-9]|100)$ ) { return 502; } - - upstream chromeai_upstream { - server 127.0.0.1:$nano_port fail_timeout=0; - } - - server { - listen 80; - server_name chromeai.local; - - add_header Access-Control-Allow-Origin $cors_origin always; - add_header Access-Control-Allow-Methods "GET, POST, OPTIONS" always; - add_header Access-Control-Allow-Headers "Content-Type, Authorization" always; - add_header Access-Control-Allow-Credentials "true" always; - add_header Content-Type application/json; - - if ($request_method = OPTIONS) { return 204; } - - location / { - proxy_pass http://chromeai_upstream$nano_path; - proxy_set_header Host $host; - proxy_set_header X-Real-IP $remote_addr; - proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for; - proxy_set_header X-Forwarded-Proto $scheme; - proxy_redirect off; - } - } -} diff --git a/panel/panel.html b/panel/panel.html index 6f4e521..92d7af5 100644 --- a/panel/panel.html +++ b/panel/panel.html @@ -157,13 +157,6 @@

Turn selected text into structure

-
- Advanced tools -
- - -
-
diff --git a/panel/panel.js b/panel/panel.js index 288e8af..2968197 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -1463,30 +1463,6 @@ function bindActions() { }); syncOrderButtonEl?.addEventListener('click', wrap(() => doSyncOrderToken())); orderIdInputEl?.addEventListener('input', () => syncControlAvailability()); - $('#btn-transcribe')?.addEventListener('click', () => wrap(async () => { - const res = await request('panel:transcribe'); - renderOutput({ - title: 'Transcription', - markdown: res.text || '', - eyebrow: 'Advanced', - meta: 'Experimental audio path.', - exportBase: 'selectpilot-transcript', - }); - renderExports({ markdown: res.text || '', basename: 'selectpilot-transcript' }); - setStatus('Done'); - })()); - $('#btn-vision')?.addEventListener('click', () => wrap(async () => { - const res = await request('panel:vision'); - renderOutput({ - title: 'Vision', - markdown: res.text || '', - eyebrow: 'Advanced', - meta: 'Experimental OCR / image signature path.', - exportBase: 'selectpilot-vision', - }); - renderExports({ markdown: res.text || '', basename: 'selectpilot-vision' }); - setStatus('Done'); - })()); extractPresetEl?.addEventListener('change', () => syncPresetHelp()); runtimeStateEl?.addEventListener('click', (event) => { const target = event.target; diff --git a/panel/panel.ts b/panel/panel.ts index 556195e..7009ddd 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -1666,34 +1666,6 @@ function bindActions() { }); syncOrderButtonEl?.addEventListener('click', wrap(() => doSyncOrderToken())); orderIdInputEl?.addEventListener('input', () => syncControlAvailability()); - $('#btn-transcribe')?.addEventListener('click', () => - wrap(async () => { - const res = await request('panel:transcribe'); - renderOutput({ - title: 'Transcription', - markdown: res.text || '', - eyebrow: 'Advanced', - meta: 'Experimental audio path.', - exportBase: 'selectpilot-transcript', - }); - renderExports({ markdown: res.text || '', basename: 'selectpilot-transcript' }); - setStatus('Done'); - })() - ); - $('#btn-vision')?.addEventListener('click', () => - wrap(async () => { - const res = await request('panel:vision'); - renderOutput({ - title: 'Vision', - markdown: res.text || '', - eyebrow: 'Advanced', - meta: 'Experimental OCR / image signature path.', - exportBase: 'selectpilot-vision', - }); - renderExports({ markdown: res.text || '', basename: 'selectpilot-vision' }); - setStatus('Done'); - })() - ); extractPresetEl?.addEventListener('change', () => syncPresetHelp()); runtimeStateEl?.addEventListener('click', (event) => { const target = event.target as HTMLElement | null; diff --git a/server/nano_server.py b/server/nano_server.py index d119e70..1cb2de0 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -43,8 +43,6 @@ "/extract", "/agent", "/embed", - "/transcribe", - "/vision", "/license/verify", ] @@ -88,18 +86,6 @@ class OperationContract: template="embed.v1", allowed_fields=("text", "session_id"), ), - "/transcribe": OperationContract( - name="transcribe", - endpoint="/transcribe", - template="transcribe.v1", - allowed_fields=("audioUrl", "mediaId", "session_id"), - ), - "/vision": OperationContract( - name="vision", - endpoint="/vision", - template="vision.v1", - allowed_fields=("imageBase64", "videoFrame", "session_id"), - ), "/license/verify": OperationContract( name="license_verify", endpoint="/license/verify", @@ -1230,19 +1216,6 @@ def build_privacy_proof(health: dict | None = None, port: int = DEFAULT_PORT) -> } -def transcribe(payload: dict) -> dict: - source = payload.get("audioUrl") or payload.get("mediaId") or "audio" - text = f"Transcribed from {source}" - return {"text": text, "confidence": 0.95} - - -def vision(payload: dict) -> dict: - blob = payload.get("imageBase64") or payload.get("videoFrame") or "" - digest = hashlib.sha256(blob.encode("utf-8")).hexdigest() if blob else "" - text = f"Image signature {digest[:16]}" if digest else "No image provided" - return {"text": text, "tags": ["ocr", "frame"] if blob else []} - - def embed(payload: dict) -> dict: text = payload.get("text", "") try: @@ -1623,10 +1596,6 @@ def _execute_summarize(): "reason": ((compiled.get("model_selection") or {}).get("reason") if isinstance(compiled.get("model_selection"), dict) else ""), }, ) - elif path == '/transcribe': - resp = transcribe(payload) - elif path == '/vision': - resp = vision(payload) elif path == '/embed': resp = embed(payload) elif path == '/agent': diff --git a/tests/e2e/panel-harness.html b/tests/e2e/panel-harness.html index 906fc77..9b166b5 100644 --- a/tests/e2e/panel-harness.html +++ b/tests/e2e/panel-harness.html @@ -95,13 +95,6 @@

Attach paid token

-
- Advanced tools -
- - -
-
diff --git a/tests/panel/capability-truth.test.mjs b/tests/panel/capability-truth.test.mjs new file mode 100644 index 0000000..5c594e5 --- /dev/null +++ b/tests/panel/capability-truth.test.mjs @@ -0,0 +1,21 @@ +// module_name: tests_panel_capability_truth_test_mjs +// spec_ref: "testing_strategy.integration_tests" +import test from 'node:test'; +import assert from 'node:assert/strict'; +import { readFile } from 'node:fs/promises'; + +import { endpoints } from '../../api/endpoints.js'; + +test('unimplemented media capabilities are absent from the shipped extension', async () => { + const [panel, harness, background] = await Promise.all([ + readFile(new URL('../../panel/panel.html', import.meta.url), 'utf8'), + readFile(new URL('../e2e/panel-harness.html', import.meta.url), 'utf8'), + readFile(new URL('../../background/background.js', import.meta.url), 'utf8'), + ]); + + assert.equal('transcribe' in endpoints, false); + assert.equal('vision' in endpoints, false); + assert.doesNotMatch(panel, /btn-(?:transcribe|vision)/); + assert.doesNotMatch(harness, /btn-(?:transcribe|vision)/); + assert.doesNotMatch(background, /panel:(?:transcribe|vision)/); +}); diff --git a/tests/server/test_capability_truth.py b/tests/server/test_capability_truth.py new file mode 100644 index 0000000..79570ec --- /dev/null +++ b/tests/server/test_capability_truth.py @@ -0,0 +1,23 @@ +# module_name: tests_server_test_capability_truth_py +# spec_ref: "testing_strategy.integration_tests" +import sys +import unittest +from pathlib import Path + + +SERVER_DIR = Path(__file__).resolve().parents[2] / "server" +sys.path.insert(0, str(SERVER_DIR)) + +import nano_server # noqa: E402 + + +class CapabilityTruthTests(unittest.TestCase): + def test_unimplemented_media_routes_are_not_admitted(self): + self.assertNotIn("/transcribe", nano_server.ALLOWED_BRIDGE_ENDPOINT_PATHS) + self.assertNotIn("/vision", nano_server.ALLOWED_BRIDGE_ENDPOINT_PATHS) + self.assertNotIn("/transcribe", nano_server.OPERATION_CONTRACTS) + self.assertNotIn("/vision", nano_server.OPERATION_CONTRACTS) + + +if __name__ == "__main__": + unittest.main() From b824386fcb8095fd12724ea813fe030261c110fb Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:12:14 +0200 Subject: [PATCH 34/66] feat(release): complete deployable SelectPilot product Add consent-driven installation, least-privilege capture, signed access, public checkout surfaces, and release automation.\n\nVerified: pnpm typecheck && pnpm lint && pnpm test && pnpm test:privacy && pnpm build && pnpm test:e2e && pnpm package:store --- .dockerignore | 5 + .github/workflows/ci.yml | 18 + .github/workflows/deploy-site.yml | 45 ++ .github/workflows/macos-helper.yml | 29 ++ .gitignore | 1 + README.md | 4 +- api/endpoints.js | 6 +- api/endpoints.ts | 7 +- background/background.js | 20 +- background/background.ts | 20 +- background/entitlement-service.js | 92 ++-- background/entitlement-service.ts | 116 +++-- content/content-script.bundle.js | 12 +- content/extract-text.js | 12 +- content/extract-text.ts | 13 +- docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md | 5 +- docs/CHROME_WEB_STORE_SUBMISSION.md | 10 +- docs/ENTITLEMENT_DEPLOYMENT.md | 33 ++ docs/PLATFORM_SUPPORT.md | 13 + docs/PRIVACY_POLICY.md | 6 +- .../com.selectpilot.bridge.plist.template | 29 ++ installer/macos/scripts/postinstall | 49 ++ launchd/com.chromeai.nano.plist | 1 + manifest.json | 23 +- options/options.css | 2 + options/options.html | 1 + options/options.js | 35 ++ options/options.ts | 46 ++ package.json | 4 + panel/panel.css | 111 ++++- panel/panel.html | 57 +-- panel/panel.js | 402 +++++++--------- panel/panel.ts | 446 +++++++----------- popup/popup.js | 10 + popup/popup.ts | 10 + pricing/entitlement-public-keys.json | 11 + scripts/build-site-config.mjs | 24 + scripts/lint-manifest.mjs | 10 +- scripts/package-chrome-store.mjs | 26 +- scripts/package-macos-helper.sh | 37 ++ scripts/verify-entitlement-key.mjs | 16 + server/installation_manager.py | 155 ++++++ server/nano_server.py | 98 +++- services/entitlement-authority/Dockerfile | 7 + services/entitlement-authority/compose.yaml | 39 ++ services/entitlement-authority/server.mjs | 285 +++++++++++ site/checkout.html | 1 + site/checkout.js | 45 ++ site/config.example.js | 6 + site/index.html | 15 + site/pricing.html | 1 + site/privacy.html | 1 + site/styles.css | 72 +++ site/support.html | 1 + tests/e2e/extension-first-run-flow.spec.mjs | 51 +- tests/e2e/panel-harness.html | 43 +- tests/e2e/panel-user-flow.spec.mjs | 17 +- tests/panel/entitlement-authority.test.mjs | 125 +++++ tests/panel/store-package.test.mjs | 35 +- tests/server/test_entitlement_verification.py | 20 +- tests/server/test_installation_manager.py | 39 ++ 61 files changed, 2099 insertions(+), 774 deletions(-) create mode 100644 .dockerignore create mode 100644 .github/workflows/deploy-site.yml create mode 100644 .github/workflows/macos-helper.yml create mode 100644 docs/ENTITLEMENT_DEPLOYMENT.md create mode 100644 docs/PLATFORM_SUPPORT.md create mode 100644 installer/macos/com.selectpilot.bridge.plist.template create mode 100644 installer/macos/scripts/postinstall create mode 100644 options/options.css create mode 100644 options/options.html create mode 100644 options/options.js create mode 100644 options/options.ts create mode 100644 pricing/entitlement-public-keys.json create mode 100644 scripts/build-site-config.mjs create mode 100644 scripts/package-macos-helper.sh create mode 100644 scripts/verify-entitlement-key.mjs create mode 100644 server/installation_manager.py create mode 100644 services/entitlement-authority/Dockerfile create mode 100644 services/entitlement-authority/compose.yaml create mode 100644 services/entitlement-authority/server.mjs create mode 100644 site/checkout.html create mode 100644 site/checkout.js create mode 100644 site/config.example.js create mode 100644 site/index.html create mode 100644 site/pricing.html create mode 100644 site/privacy.html create mode 100644 site/styles.css create mode 100644 site/support.html create mode 100644 tests/panel/entitlement-authority.test.mjs create mode 100644 tests/server/test_installation_manager.py diff --git a/.dockerignore b/.dockerignore new file mode 100644 index 0000000..767e27b --- /dev/null +++ b/.dockerignore @@ -0,0 +1,5 @@ +** +!services/ +!services/entitlement-authority/ +!services/entitlement-authority/server.mjs +!services/entitlement-authority/Dockerfile diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 3ec62c2..5514067 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -73,6 +73,9 @@ jobs: - name: Validate Chrome Web Store assets run: pnpm validate:store + - name: Package Chrome Web Store release candidate + run: pnpm package:store + - name: Smoke direct server entrypoint run: python3 server/nano_server.py --help @@ -115,3 +118,18 @@ jobs: path: reports/code_admission_evidence.json if-no-files-found: error retention-days: 30 + + - name: Upload Chrome Web Store release candidate + if: matrix.node-version == '22.x' + uses: actions/upload-artifact@v7 + with: + name: selectpilot-chrome-web-store-${{ github.sha }} + path: | + dist/chrome-web-store/*.zip + dist/chrome-web-store/*.zip.json + if-no-files-found: error + retention-days: 30 + + - name: Build entitlement authority image + if: matrix.node-version == '22.x' + run: docker build -f services/entitlement-authority/Dockerfile -t selectpilot-entitlement:${{ github.sha }} . diff --git a/.github/workflows/deploy-site.yml b/.github/workflows/deploy-site.yml new file mode 100644 index 0000000..a940822 --- /dev/null +++ b/.github/workflows/deploy-site.yml @@ -0,0 +1,45 @@ +name: Deploy public site + +on: + workflow_dispatch: + push: + branches: [main] + paths: ["site/**", "scripts/build-site-config.mjs", ".github/workflows/deploy-site.yml"] + +permissions: + contents: read + pages: write + id-token: write + +concurrency: + group: pages + cancel-in-progress: false + +jobs: + deploy: + environment: + name: github-pages + url: ${{ steps.deployment.outputs.page_url }} + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v7 + - uses: actions/setup-node@v7 + with: + node-version: 22 + - name: Build configured public site + run: node scripts/build-site-config.mjs + env: + PADDLE_ENVIRONMENT: production + PADDLE_CLIENT_TOKEN: ${{ secrets.PADDLE_CLIENT_TOKEN }} + PADDLE_PRICE_ESSENTIAL: ${{ vars.PADDLE_PRICE_ESSENTIAL }} + PADDLE_PRICE_PLUS: ${{ vars.PADDLE_PRICE_PLUS }} + PADDLE_PRICE_PRO: ${{ vars.PADDLE_PRICE_PRO }} + ENTITLEMENT_AUTHORITY_URL: ${{ vars.ENTITLEMENT_AUTHORITY_URL }} + CHROME_WEB_STORE_URL: ${{ vars.CHROME_WEB_STORE_URL }} + - uses: actions/configure-pages@v5 + - uses: actions/upload-pages-artifact@v4 + with: + path: site-dist + - name: Deploy + id: deployment + uses: actions/deploy-pages@v4 diff --git a/.github/workflows/macos-helper.yml b/.github/workflows/macos-helper.yml new file mode 100644 index 0000000..461ba7b --- /dev/null +++ b/.github/workflows/macos-helper.yml @@ -0,0 +1,29 @@ +name: macOS helper + +on: + workflow_dispatch: + pull_request: + branches: [main] + paths: ["installer/macos/**", "server/**", "scripts/package-macos-helper.sh", ".github/workflows/macos-helper.yml"] + +permissions: + contents: read + +jobs: + package: + runs-on: macos-15 + steps: + - uses: actions/checkout@v7 + - uses: actions/setup-node@v7 + with: + node-version: 22 + - name: Package helper release candidate + run: sh scripts/package-macos-helper.sh + - name: Inspect package + run: pkgutil --payload-files dist/macos-helper/SelectPilot-Installer-unsigned.pkg | sort + - uses: actions/upload-artifact@v7 + with: + name: selectpilot-macos-helper-unsigned-${{ github.sha }} + path: dist/macos-helper/SelectPilot-Installer-unsigned.pkg + if-no-files-found: error + retention-days: 14 diff --git a/.gitignore b/.gitignore index 2da88e8..6360eef 100644 --- a/.gitignore +++ b/.gitignore @@ -7,5 +7,6 @@ test-results/ playwright-report/ reports/code_admission_evidence.json dist/ +site-dist/ /usr/local/var/log/chromeai/ /usr/local/var/run/chromeai/ diff --git a/README.md b/README.md index 4d45ff7..239c915 100644 --- a/README.md +++ b/README.md @@ -125,7 +125,7 @@ The repository configuration is authoritative for price and entitlement mapping: Essential and Plus do not retain a knowledge history. Pro stateful features must remain local, visible and user-controlled: retained data can be inspected, exported and deleted. A feature is available only when the signed entitlement and `pricing/tier-feature-map.json` both permit it; there is no silent downgrade or fallback. -Displayed tier prices live in `pricing/pricing-global.json`. Checkout is not embedded in the extension: paid tokens come from the separately operated local entitlement authority, and Store packaging excludes inactive remote-checkout code and product placeholders. +Displayed prices live in `pricing/pricing-global.json`. The extension opens SelectPilot's HTTPS pricing page; Paddle and its deployment-injected price IDs run only on that public site. Store packaging excludes checkout code and product identifiers from the extension runtime. ### Team / self-hosted @@ -214,7 +214,7 @@ Store listing copy, privacy language, and the release checklist live in: - [`docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md`](docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md) - [`docs/PRIVACY_POLICY.md`](docs/PRIVACY_POLICY.md) -`pnpm package:store` builds and audits the upload ZIP. It intentionally fails closed until production entitlement signature verification is configured; no unsigned commercial package is release-eligible. +`pnpm package:store` builds and audits the upload ZIP. It fails closed unless the production entitlement public key is pinned and valid; unsigned or remotely executed commercial logic is never release-eligible. ## Compute Distribution diff --git a/api/endpoints.js b/api/endpoints.js index 136d323..7250a19 100644 --- a/api/endpoints.js +++ b/api/endpoints.js @@ -1,7 +1,6 @@ // module_name: api_endpoints_ts // spec_ref: "execution_layer" export const API_BASE = 'http://127.0.0.1:8083'; -export const BILLING_BASE = 'http://127.0.0.1:8090'; export const endpoints = { health: `${API_BASE}/health`, profiles: `${API_BASE}/profiles`, @@ -9,11 +8,14 @@ export const endpoints = { runtimeMetaStream: `${API_BASE}/runtime-meta/stream`, intentCompile: `${API_BASE}/intent/compile`, benchmark: `${API_BASE}/benchmark`, + installationStatus: `${API_BASE}/installation/status`, + installationStart: `${API_BASE}/installation/start`, privacyProof: `${API_BASE}/privacy-proof`, summarize: `${API_BASE}/summarize`, extract: `${API_BASE}/extract`, embed: `${API_BASE}/embed`, agent: `${API_BASE}/agent`, licenseVerify: `${API_BASE}/license/verify`, - billingOrderStatus: (orderId) => `${BILLING_BASE}/order/${encodeURIComponent(orderId)}`, + licenseTrial: `${API_BASE}/license/trial`, + licenseClaim: `${API_BASE}/license/claim`, }; diff --git a/api/endpoints.ts b/api/endpoints.ts index 4e2ed07..160cc6b 100644 --- a/api/endpoints.ts +++ b/api/endpoints.ts @@ -1,8 +1,6 @@ // module_name: api_endpoints_ts // spec_ref: "execution_layer" export const API_BASE = 'http://127.0.0.1:8083'; -export const BILLING_BASE = 'http://127.0.0.1:8090'; - export const endpoints = { health: `${API_BASE}/health`, profiles: `${API_BASE}/profiles`, @@ -10,11 +8,14 @@ export const endpoints = { runtimeMetaStream: `${API_BASE}/runtime-meta/stream`, intentCompile: `${API_BASE}/intent/compile`, benchmark: `${API_BASE}/benchmark`, + installationStatus: `${API_BASE}/installation/status`, + installationStart: `${API_BASE}/installation/start`, privacyProof: `${API_BASE}/privacy-proof`, summarize: `${API_BASE}/summarize`, extract: `${API_BASE}/extract`, embed: `${API_BASE}/embed`, agent: `${API_BASE}/agent`, licenseVerify: `${API_BASE}/license/verify`, - billingOrderStatus: (orderId: string) => `${BILLING_BASE}/order/${encodeURIComponent(orderId)}`, + licenseTrial: `${API_BASE}/license/trial`, + licenseClaim: `${API_BASE}/license/claim`, }; diff --git a/background/background.js b/background/background.js index 63dba21..701f763 100644 --- a/background/background.js +++ b/background/background.js @@ -9,6 +9,21 @@ import { ApiRequestError } from '../api/request.js'; import { FIRST_RUN_EXAMPLE } from '../shared/first-run-example.js'; const MEMORY_ENABLED_KEY = 'selectpilot_memory_enabled_v1'; const MEMORY_LEDGER_KEY = 'selectpilot_memory_ledger_v1'; +async function openForActiveTab() { + const [tab] = await chrome.tabs.query({ active: true, currentWindow: true }); + if (!tab?.id || tab.windowId === undefined) + throw new Error('active_tab_unavailable'); + await chrome.scripting.executeScript({ + target: { tabId: tab.id }, + files: ['content/content-script.bundle.js'], + }); + await chrome.sidePanel.setOptions({ tabId: tab.id, path: 'panel/panel.html', enabled: true }); + await chrome.sidePanel.open({ windowId: tab.windowId }); +} +chrome.commands.onCommand.addListener((command) => { + if (command === 'open-selectpilot') + void openForActiveTab().catch((cause) => error('shortcut', 'open failed', cause)); +}); async function canUseProjectMemory() { const { allowed } = await requireFeature('project_memory'); return allowed; @@ -130,6 +145,7 @@ async function collectContext() { const documentText = doc?.documentText?.text || ''; const url = selection?.text?.url || doc?.documentText?.url || tab.url; const title = selection?.text?.title || doc?.documentText?.title || tab.title; + const pageColor = selection?.text?.pageColor || doc?.documentText?.pageColor; return { url: url || undefined, title: title || undefined, @@ -143,7 +159,8 @@ async function collectContext() { metadata: { audioDuration: audio?.audio?.duration, videoDuration: video?.video?.duration, - capturedAt: Date.now() + capturedAt: Date.now(), + pageColor, } }; } @@ -241,6 +258,7 @@ async function handleSelectionPreview() { title: context.title || '', url: context.url || '', hasSelection: Boolean(context.selection && context.selection.trim()), + pageColor: typeof context.metadata?.pageColor === 'string' ? context.metadata.pageColor : '', }; } async function handleIntentCompile(intent) { diff --git a/background/background.ts b/background/background.ts index be4d3b7..1c4d6ed 100644 --- a/background/background.ts +++ b/background/background.ts @@ -27,6 +27,21 @@ type MemoryEntry = { const MEMORY_ENABLED_KEY = 'selectpilot_memory_enabled_v1'; const MEMORY_LEDGER_KEY = 'selectpilot_memory_ledger_v1'; +async function openForActiveTab(): Promise { + const [tab] = await chrome.tabs.query({ active: true, currentWindow: true }); + if (!tab?.id || tab.windowId === undefined) throw new Error('active_tab_unavailable'); + await chrome.scripting.executeScript({ + target: { tabId: tab.id }, + files: ['content/content-script.bundle.js'], + }); + await chrome.sidePanel.setOptions({ tabId: tab.id, path: 'panel/panel.html', enabled: true }); + await chrome.sidePanel.open({ windowId: tab.windowId }); +} + +chrome.commands.onCommand.addListener((command) => { + if (command === 'open-selectpilot') void openForActiveTab().catch((cause) => error('shortcut', 'open failed', cause)); +}); + async function canUseProjectMemory(): Promise { const { allowed } = await requireFeature('project_memory'); return allowed; @@ -156,6 +171,7 @@ async function collectContext(): Promise { const documentText = (doc as any)?.documentText?.text || ''; const url = (selection as any)?.text?.url || (doc as any)?.documentText?.url || tab.url; const title = (selection as any)?.text?.title || (doc as any)?.documentText?.title || tab.title; + const pageColor = (selection as any)?.text?.pageColor || (doc as any)?.documentText?.pageColor; return { url: url || undefined, @@ -170,7 +186,8 @@ async function collectContext(): Promise { metadata: { audioDuration: (audio as any)?.audio?.duration, videoDuration: (video as any)?.video?.duration, - capturedAt: Date.now() + capturedAt: Date.now(), + pageColor, } }; } @@ -270,6 +287,7 @@ async function handleSelectionPreview(): Promise { title: context.title || '', url: context.url || '', hasSelection: Boolean(context.selection && context.selection.trim()), + pageColor: typeof context.metadata?.pageColor === 'string' ? context.metadata.pageColor : '', }; } diff --git a/background/entitlement-service.js b/background/entitlement-service.js index d258059..671acae 100644 --- a/background/entitlement-service.js +++ b/background/entitlement-service.js @@ -5,26 +5,14 @@ import { error, log, warn } from '../utils/logger.js'; import { loadLicense, loadToken, saveLicense } from '../licensing/license-storage.js'; const OFFLINE_GRACE_MS = 7 * 24 * 60 * 60 * 1000; const REFRESH_INTERVAL_MS = 10 * 60 * 1000; -const ENTITLEMENT_PUBLIC_KEYS = { - __SELECTPILOT_ENTITLEMENT_KEY_ID__: '__SELECTPILOT_ENTITLEMENT_PUBLIC_KEY_HEX__', -}; const SIGNATURE_ALGORITHM = 'Ed25519'; let cachedFeatureMap = null; +let cachedKeyring = null; function nowMs() { return Date.now(); } -function normalizeEntitlement(payload) { - return { - token: payload.token, - tier: payload.tier, - issuedAt: payload.issuedAt, - expiresAt: payload.expiresAt ?? undefined, - features: payload.features, - cachedAt: nowMs(), - }; -} function hexToBytes(hex) { - if (!hex || hex.length % 2 !== 0) + if (!/^[0-9a-f]{64}$/i.test(hex)) throw new Error('invalid public key hex'); const bytes = new Uint8Array(hex.length / 2); for (let i = 0; i < bytes.length; i++) { @@ -53,12 +41,35 @@ function canonicalizeEntitlement(payload) { expiresAt: payload.expiresAt ?? null, }); } -export async function verifyEntitlementSignature(payload, signature, kid, publicKeys = ENTITLEMENT_PUBLIC_KEYS) { - const publicKeyHex = publicKeys[kid]; - if (!publicKeyHex || !signature || !kid) +async function loadKeyring() { + if (cachedKeyring) + return cachedKeyring; + const url = chrome.runtime.getURL('pricing/entitlement-public-keys.json'); + const res = await fetch(url, { cache: 'no-store' }); + if (!res.ok) + throw new Error('entitlement_keyring_unavailable'); + const value = (await res.json()); + if (value.schema_version !== 1 || !Array.isArray(value.keys) || value.keys.length === 0) { + throw new Error('entitlement_keyring_invalid'); + } + if (value.keys.some((key) => (!key.kid + || key.alg !== 'Ed25519' + || !/^[0-9a-f]{64}$/i.test(key.public_key_hex) + || !['active', 'retiring'].includes(key.status)))) + throw new Error('entitlement_keyring_invalid'); + cachedKeyring = value; + return value; +} +export async function verifyEntitlementSignature(payload, signature, kid, publicKeys) { + if (!kid || !signature) return false; try { - const key = await crypto.subtle.importKey('raw', bytesToArrayBuffer(hexToBytes(publicKeyHex)), { name: 'Ed25519' }, false, ['verify']); + const keyEntry = publicKeys + ? { public_key_hex: publicKeys[kid] } + : (await loadKeyring()).keys.find((entry) => (entry.kid === kid && entry.alg === SIGNATURE_ALGORITHM && ['active', 'retiring'].includes(entry.status))); + if (!keyEntry?.public_key_hex) + return false; + const key = await crypto.subtle.importKey('raw', bytesToArrayBuffer(hexToBytes(keyEntry.public_key_hex)), { name: 'Ed25519' }, false, ['verify']); const valid = await crypto.subtle.verify('Ed25519', key, bytesToArrayBuffer(base64ToBytes(signature)), new TextEncoder().encode(canonicalizeEntitlement(payload))); return Boolean(valid); } @@ -107,12 +118,23 @@ async function isVerifiedCachedEntitlement(record, token) { } async function readCachedEntitlement() { const cached = await loadLicense(); - if (!cached) + return cached ? validateCachedEntitlement(cached) : null; +} +async function validateCachedEntitlement(record) { + if (!record.signature || !record.kid || record.alg !== 'Ed25519') return null; - return { - ...cached, - cachedAt: cached.cachedAt || cached.issuedAt, + if (record.expiresAt !== undefined && nowMs() >= record.expiresAt) + return null; + const payload = { + token: record.token, + tier: record.tier, + features: record.features, + issuedAt: record.issuedAt, + expiresAt: record.expiresAt ?? null, }; + if (!(await verifyEntitlementSignature(payload, record.signature, record.kid))) + return null; + return { ...record, cachedAt: record.cachedAt || record.issuedAt }; } async function writeCachedEntitlement(record) { await saveLicense({ @@ -127,27 +149,34 @@ async function writeCachedEntitlement(record) { kid: record.kid, }); } -async function normalizeRemoteResponse(token, response) { - const entitlement = response.entitlement; +export async function normalizeRemoteResponse(token, response, keyring) { + const entitlement = response?.entitlement; if (!entitlement || entitlement.token !== token || !['essential', 'plus', 'pro'].includes(entitlement.tier) || !Number.isInteger(entitlement.issuedAt) - || (entitlement.expiresAt != null && !Number.isInteger(entitlement.expiresAt))) { - warn('entitlement', 'missing or token-mismatched signed entitlement'); + || (entitlement.expiresAt !== undefined + && entitlement.expiresAt !== null + && !Number.isInteger(entitlement.expiresAt)) + || !Array.isArray(entitlement.features) + || entitlement.features.some((feature) => typeof feature !== 'string')) { + warn('entitlement', 'invalid signed entitlement contract'); return null; } if (response.alg !== SIGNATURE_ALGORITHM || !response.signature || !response.kid) { warn('entitlement', 'unsigned or unsupported entitlement response'); return null; } - if (entitlement.expiresAt != null && entitlement.expiresAt <= nowMs()) { + if (entitlement.expiresAt !== undefined && entitlement.expiresAt !== null && nowMs() >= entitlement.expiresAt) { warn('entitlement', 'expired entitlement response'); return null; } - const valid = await verifyEntitlementSignature(entitlement, response.signature, response.kid); - if (!valid) + const publicKeys = keyring + ? Object.fromEntries(keyring.keys.map((entry) => [entry.kid, entry.public_key_hex])) + : undefined; + if (!(await verifyEntitlementSignature(entitlement, response.signature, response.kid, publicKeys))) { return null; + } return { token: entitlement.token, tier: entitlement.tier, @@ -196,7 +225,8 @@ export async function setEntitlementToken(token) { } export async function refreshEntitlement(force = false) { const token = await loadToken(); - const cached = await loadLicense(); + const stored = await loadLicense(); + const cached = stored ? await validateCachedEntitlement(stored) : null; if (!token) return null; const shouldAttemptRemote = force @@ -221,7 +251,7 @@ export async function refreshEntitlement(force = false) { } if (remote) { await writeCachedEntitlement(remote); - return normalizeEntitlement(remote); + return remote; } if (cached && await isVerifiedCachedEntitlement(cached, token)) { log('entitlement', 'remote unavailable; using cached entitlement within grace window'); diff --git a/background/entitlement-service.ts b/background/entitlement-service.ts index 0fcfe66..c7f7304 100644 --- a/background/entitlement-service.ts +++ b/background/entitlement-service.ts @@ -15,14 +15,15 @@ type EntitlementPayload = { }; type SignedEntitlementResponse = { - entitlement?: EntitlementPayload; - signature?: string; - alg?: string; - kid?: string; - token?: string; - tier?: EntitlementTier; - issuedAt?: number; - expiresAt?: number; + entitlement: EntitlementPayload; + signature: string; + alg: 'Ed25519'; + kid: string; +}; + +export type EntitlementKeyring = { + schema_version: 1; + keys: Array<{ kid: string; alg: 'Ed25519'; public_key_hex: string; status: 'active' | 'retiring' }>; }; type RemoteVerifyResult = CachedEntitlement | 'unauthorized' | null; @@ -41,30 +42,16 @@ export type CachedEntitlement = { const OFFLINE_GRACE_MS = 7 * 24 * 60 * 60 * 1000; const REFRESH_INTERVAL_MS = 10 * 60 * 1000; -const ENTITLEMENT_PUBLIC_KEYS: Readonly> = { - __SELECTPILOT_ENTITLEMENT_KEY_ID__: '__SELECTPILOT_ENTITLEMENT_PUBLIC_KEY_HEX__', -}; const SIGNATURE_ALGORITHM = 'Ed25519'; - let cachedFeatureMap: Record | null = null; +let cachedKeyring: EntitlementKeyring | null = null; function nowMs(): number { return Date.now(); } -function normalizeEntitlement(payload: EntitlementPayload): LicenseRecord { - return { - token: payload.token, - tier: payload.tier, - issuedAt: payload.issuedAt, - expiresAt: payload.expiresAt ?? undefined, - features: payload.features, - cachedAt: nowMs(), - }; -} - function hexToBytes(hex: string): Uint8Array { - if (!hex || hex.length % 2 !== 0) throw new Error('invalid public key hex'); + if (!/^[0-9a-f]{64}$/i.test(hex)) throw new Error('invalid public key hex'); const bytes = new Uint8Array(hex.length / 2); for (let i = 0; i < bytes.length; i++) { bytes[i] = parseInt(hex.slice(i * 2, i * 2 + 2), 16); @@ -95,18 +82,42 @@ function canonicalizeEntitlement(payload: EntitlementPayload): string { }); } +async function loadKeyring(): Promise { + if (cachedKeyring) return cachedKeyring; + const url = chrome.runtime.getURL('pricing/entitlement-public-keys.json'); + const res = await fetch(url, { cache: 'no-store' }); + if (!res.ok) throw new Error('entitlement_keyring_unavailable'); + const value = (await res.json()) as EntitlementKeyring; + if (value.schema_version !== 1 || !Array.isArray(value.keys) || value.keys.length === 0) { + throw new Error('entitlement_keyring_invalid'); + } + if (value.keys.some((key) => ( + !key.kid + || key.alg !== 'Ed25519' + || !/^[0-9a-f]{64}$/i.test(key.public_key_hex) + || !['active', 'retiring'].includes(key.status) + ))) throw new Error('entitlement_keyring_invalid'); + cachedKeyring = value; + return value; +} + export async function verifyEntitlementSignature( payload: EntitlementPayload, signature: string, kid: string, - publicKeys: Readonly> = ENTITLEMENT_PUBLIC_KEYS, + publicKeys?: Readonly>, ): Promise { - const publicKeyHex = publicKeys[kid]; - if (!publicKeyHex || !signature || !kid) return false; + if (!kid || !signature) return false; try { + const keyEntry = publicKeys + ? { public_key_hex: publicKeys[kid] } + : (await loadKeyring()).keys.find((entry) => ( + entry.kid === kid && entry.alg === SIGNATURE_ALGORITHM && ['active', 'retiring'].includes(entry.status) + )); + if (!keyEntry?.public_key_hex) return false; const key = await crypto.subtle.importKey( 'raw', - bytesToArrayBuffer(hexToBytes(publicKeyHex)), + bytesToArrayBuffer(hexToBytes(keyEntry.public_key_hex)), { name: 'Ed25519' }, false, ['verify'] @@ -167,11 +178,21 @@ async function isVerifiedCachedEntitlement(record: LicenseRecord, token: string) async function readCachedEntitlement(): Promise { const cached = await loadLicense(); - if (!cached) return null; - return { - ...cached, - cachedAt: cached.cachedAt || cached.issuedAt, + return cached ? validateCachedEntitlement(cached) : null; +} + +async function validateCachedEntitlement(record: LicenseRecord): Promise { + if (!record.signature || !record.kid || record.alg !== 'Ed25519') return null; + if (record.expiresAt !== undefined && nowMs() >= record.expiresAt) return null; + const payload: EntitlementPayload = { + token: record.token, + tier: record.tier, + features: record.features, + issuedAt: record.issuedAt, + expiresAt: record.expiresAt ?? null, }; + if (!(await verifyEntitlementSignature(payload, record.signature, record.kid))) return null; + return { ...record, cachedAt: record.cachedAt || record.issuedAt } as CachedEntitlement; } async function writeCachedEntitlement(record: CachedEntitlement): Promise { @@ -188,28 +209,40 @@ async function writeCachedEntitlement(record: CachedEntitlement): Promise }); } -async function normalizeRemoteResponse(token: string, response: SignedEntitlementResponse): Promise { - const entitlement = response.entitlement; +export async function normalizeRemoteResponse( + token: string, + response: SignedEntitlementResponse, + keyring?: EntitlementKeyring +): Promise { + const entitlement = response?.entitlement; if ( !entitlement || entitlement.token !== token || !['essential', 'plus', 'pro'].includes(entitlement.tier) || !Number.isInteger(entitlement.issuedAt) - || (entitlement.expiresAt != null && !Number.isInteger(entitlement.expiresAt)) + || (entitlement.expiresAt !== undefined + && entitlement.expiresAt !== null + && !Number.isInteger(entitlement.expiresAt)) + || !Array.isArray(entitlement.features) + || entitlement.features.some((feature) => typeof feature !== 'string') ) { - warn('entitlement', 'missing or token-mismatched signed entitlement'); + warn('entitlement', 'invalid signed entitlement contract'); return null; } if (response.alg !== SIGNATURE_ALGORITHM || !response.signature || !response.kid) { warn('entitlement', 'unsigned or unsupported entitlement response'); return null; } - if (entitlement.expiresAt != null && entitlement.expiresAt <= nowMs()) { + if (entitlement.expiresAt !== undefined && entitlement.expiresAt !== null && nowMs() >= entitlement.expiresAt) { warn('entitlement', 'expired entitlement response'); return null; } - const valid = await verifyEntitlementSignature(entitlement, response.signature, response.kid); - if (!valid) return null; + const publicKeys = keyring + ? Object.fromEntries(keyring.keys.map((entry) => [entry.kid, entry.public_key_hex])) + : undefined; + if (!(await verifyEntitlementSignature(entitlement, response.signature, response.kid, publicKeys))) { + return null; + } return { token: entitlement.token, tier: entitlement.tier, @@ -260,7 +293,8 @@ export async function setEntitlementToken(token: string): Promise { export async function refreshEntitlement(force = false): Promise { const token = await loadToken(); - const cached = await loadLicense(); + const stored = await loadLicense(); + const cached = stored ? await validateCachedEntitlement(stored) : null; if (!token) return null; @@ -288,7 +322,7 @@ export async function refreshEntitlement(force = false): Promise + + + Labelcom.selectpilot.bridge + ProgramArguments + /usr/bin/python3__PROJECT_ROOT__/server/nano_server.py + --port8083 + --run-dir__RUN_DIR__ + --log-dir__LOG_DIR__ + --binary-path__PROJECT_ROOT__/server/nano_server.py + + RunAtLoad + KeepAliveSuccessfulExit + StandardOutPath__LOG_DIR__/bridge.log + StandardErrorPath__LOG_DIR__/bridge.err + WorkingDirectory__PROJECT_ROOT__/server + EnvironmentVariables + CHROMEAI_PORT_INFO__RUN_DIR__/port.info + CHROMEAI_EXPECTED_PORT_FILE__RUN_DIR__/port.info + CHROMEAI_ALLOWED_RANGE8083 + CHROMEAI_OLLAMA_BASE_URL__OLLAMA_BASE_URL__ + CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ + CHROMEAI_OLLAMA_EMBED_MODEL__OLLAMA_EMBED_MODEL__ + CHROMEAI_OLLAMA_NUM_CTX__OLLAMA_NUM_CTX__ + CHROMEAI_BINARY_HASH__BINARY_HASH__ + CHROMEAI_AUTO_PULL_MODELS0 + SELECTPILOT_ENTITLEMENT_AUTHORITY_URL__ENTITLEMENT_AUTHORITY_URL__ + + diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall new file mode 100644 index 0000000..7eeba2e --- /dev/null +++ b/installer/macos/scripts/postinstall @@ -0,0 +1,49 @@ +#!/bin/sh +set -eu + +CONSOLE_USER="$(stat -f '%Su' /dev/console)" +if [ -z "$CONSOLE_USER" ] || [ "$CONSOLE_USER" = "root" ]; then + echo "A logged-in macOS user is required." >&2 + exit 1 +fi + +USER_HOME="$(dscl . -read "/Users/$CONSOLE_USER" NFSHomeDirectory | awk '{print $2}')" +USER_ID="$(id -u "$CONSOLE_USER")" +SOURCE_ROOT="/Library/Application Support/SelectPilot" +RUN_DIR="$USER_HOME/Library/Application Support/SelectPilot/run" +LOG_DIR="$USER_HOME/Library/Logs/SelectPilot" +AGENT_DIR="$USER_HOME/Library/LaunchAgents" +AGENT_PATH="$AGENT_DIR/com.selectpilot.bridge.plist" +HASH="$(shasum -a 256 "$SOURCE_ROOT/server/nano_server.py" | awk '{print $1}')" +PROFILE_VALUES="$(sudo -u "$CONSOLE_USER" /usr/bin/python3 - "$SOURCE_ROOT/server" <<'PY' +import sys +sys.path.insert(0, sys.argv[1]) +from runtime_profiles import get_runtime_profile, recommend_runtime_profile +profile = get_runtime_profile(recommend_runtime_profile()["recommended_profile"]) +print(f"{profile.generation_model}|{profile.embedding_model}|{profile.num_ctx}") +PY +)" +IFS='|' read -r OLLAMA_MODEL OLLAMA_EMBED_MODEL OLLAMA_NUM_CTX < "$AGENT_PATH" +chown "$CONSOLE_USER" "$AGENT_PATH" +chmod 600 "$AGENT_PATH" + +launchctl bootout "gui/$USER_ID" "$AGENT_PATH" >/dev/null 2>&1 || true +launchctl bootstrap "gui/$USER_ID" "$AGENT_PATH" +launchctl enable "gui/$USER_ID/com.selectpilot.bridge" + +exit 0 diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index 56b69c5..13f4506 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -27,6 +27,7 @@ CHROMEAI_OLLAMA_EMBED_MODEL__OLLAMA_EMBED_MODEL__ CHROMEAI_OLLAMA_NUM_CTX__OLLAMA_NUM_CTX__ CHROMEAI_BINARY_HASH__BINARY_HASH__ + CHROMEAI_AUTO_PULL_MODELS0 diff --git a/manifest.json b/manifest.json index d26651d..5737354 100644 --- a/manifest.json +++ b/manifest.json @@ -12,7 +12,7 @@ "256": "assets/icon256.png", "512": "assets/icon512.png" }, - "permissions": ["activeTab", "storage", "scripting", "tabs", "sidePanel"], + "permissions": ["activeTab", "storage", "scripting", "sidePanel"], "host_permissions": ["http://127.0.0.1/*", "http://localhost/*"], "background": { "service_worker": "background/background.js", @@ -30,17 +30,14 @@ "side_panel": { "default_path": "panel/panel.html" }, - "content_scripts": [ - { - "matches": [""], - "js": ["content/content-script.bundle.js"], - "run_at": "document_idle" - } - ], - "web_accessible_resources": [ - { - "resources": ["assets/icon16.png", "assets/icon32.png", "assets/icon48.png", "assets/icon128.png", "assets/icon256.png", "assets/icon512.png"], - "matches": [""] + "options_ui": { + "page": "options/options.html", + "open_in_tab": true + }, + "commands": { + "open-selectpilot": { + "suggested_key": { "default": "Alt+Shift+S", "mac": "Command+Shift+S" }, + "description": "Open SelectPilot for the current selection" } - ] + } } diff --git a/options/options.css b/options/options.css new file mode 100644 index 0000000..98ca3fe --- /dev/null +++ b/options/options.css @@ -0,0 +1,2 @@ +:root { --ink:#17201d; --muted:#65716d; --paper:#eef2ed; --glass:rgba(255,255,255,.68); --line:rgba(23,32,29,.12); --moss:#48695d; } +*{box-sizing:border-box}body{margin:0;min-height:100vh;color:var(--ink);background:radial-gradient(circle at 15% 0,rgba(141,184,173,.4),transparent 34rem),linear-gradient(150deg,#fafbf7,var(--paper));font:15px/1.5 "Avenir Next","Helvetica Neue",sans-serif}main{width:min(720px,calc(100% - 32px));margin:48px auto;display:grid;gap:16px}header,section,footer{border:1px solid var(--line);background:var(--glass);backdrop-filter:blur(24px);border-radius:28px 36px 31px 40px;padding:26px}header{display:flex;gap:18px;align-items:center}.mark{width:48px;height:48px;border-radius:16px 22px 17px 24px;background:linear-gradient(145deg,var(--moss),#8db8ad)}header span{color:var(--moss);font-size:11px;letter-spacing:.14em;text-transform:uppercase}h1{margin:4px 0 0;font:500 34px/1.1 Georgia,serif;letter-spacing:-.03em}h2{margin:0 0 18px;font-size:16px}label{display:flex;align-items:center;justify-content:space-between;gap:20px;padding:14px 0;border-top:1px solid var(--line)}label:first-of-type{border-top:0}select,button{font:inherit;border:1px solid var(--line);border-radius:14px 18px 15px 20px;background:rgba(255,255,255,.8);padding:10px 14px;color:var(--ink)}input{accent-color:var(--moss);width:20px;height:20px}.note,footer{color:var(--muted);font-size:13px}.actions{display:flex;flex-wrap:wrap;gap:10px}button{cursor:pointer}@media(max-width:520px){main{margin:16px auto}header,section,footer{padding:20px}label{align-items:flex-start;flex-direction:column}} diff --git a/options/options.html b/options/options.html new file mode 100644 index 0000000..0f0b72e --- /dev/null +++ b/options/options.html @@ -0,0 +1 @@ +SelectPilot settings
Settings

Make SelectPilot feel natural.

Flow

Chrome controls whether its side panel appears on the left or right. SelectPilot adapts its controls within that space.

Accessibility

Chrome controls

These settings belong to Chrome and open in Chrome's own interface.

Settings are saved automatically.
diff --git a/options/options.js b/options/options.js new file mode 100644 index 0000000..791c83e --- /dev/null +++ b/options/options.js @@ -0,0 +1,35 @@ +const defaults = { controlSide: 'right', resultView: 'readable', reduceMotion: false, highContrast: false, textSize: 'standard' }; +const fields = { + controlSide: document.querySelector('#control-side'), + resultView: document.querySelector('#result-view'), + reduceMotion: document.querySelector('#reduce-motion'), + highContrast: document.querySelector('#high-contrast'), + textSize: document.querySelector('#text-size'), +}; +const statusMessage = document.querySelector('#status'); +async function load() { + const stored = await chrome.storage.sync.get('selectpilot_settings'); + const settings = { ...defaults, ...(stored.selectpilot_settings || {}) }; + fields.controlSide.value = settings.controlSide; + fields.resultView.value = settings.resultView; + fields.reduceMotion.checked = settings.reduceMotion; + fields.highContrast.checked = settings.highContrast; + fields.textSize.value = settings.textSize; +} +async function save() { + const settings = { + controlSide: fields.controlSide.value, + resultView: fields.resultView.value, + reduceMotion: fields.reduceMotion.checked, + highContrast: fields.highContrast.checked, + textSize: fields.textSize.value, + }; + await chrome.storage.sync.set({ selectpilot_settings: settings }); + if (statusMessage) + statusMessage.textContent = 'Saved.'; +} +Object.values(fields).forEach((field) => field?.addEventListener('change', () => void save())); +document.querySelector('#shortcuts')?.addEventListener('click', () => void chrome.tabs.create({ url: 'chrome://extensions/shortcuts' })); +document.querySelector('#downloads')?.addEventListener('click', () => void chrome.tabs.create({ url: 'chrome://settings/downloads' })); +void load(); +export {}; diff --git a/options/options.ts b/options/options.ts new file mode 100644 index 0000000..456b6c1 --- /dev/null +++ b/options/options.ts @@ -0,0 +1,46 @@ +export {}; + +type Settings = { + controlSide: 'left' | 'right'; + resultView: 'readable' | 'structured'; + reduceMotion: boolean; + highContrast: boolean; + textSize: 'standard' | 'large' | 'larger'; +}; + +const defaults: Settings = { controlSide: 'right', resultView: 'readable', reduceMotion: false, highContrast: false, textSize: 'standard' }; +const fields = { + controlSide: document.querySelector('#control-side'), + resultView: document.querySelector('#result-view'), + reduceMotion: document.querySelector('#reduce-motion'), + highContrast: document.querySelector('#high-contrast'), + textSize: document.querySelector('#text-size'), +}; +const statusMessage = document.querySelector('#status'); + +async function load() { + const stored = await chrome.storage.sync.get('selectpilot_settings'); + const settings = { ...defaults, ...(stored.selectpilot_settings || {}) } as Settings; + fields.controlSide!.value = settings.controlSide; + fields.resultView!.value = settings.resultView; + fields.reduceMotion!.checked = settings.reduceMotion; + fields.highContrast!.checked = settings.highContrast; + fields.textSize!.value = settings.textSize; +} + +async function save() { + const settings: Settings = { + controlSide: fields.controlSide!.value as Settings['controlSide'], + resultView: fields.resultView!.value as Settings['resultView'], + reduceMotion: fields.reduceMotion!.checked, + highContrast: fields.highContrast!.checked, + textSize: fields.textSize!.value as Settings['textSize'], + }; + await chrome.storage.sync.set({ selectpilot_settings: settings }); + if (statusMessage) statusMessage.textContent = 'Saved.'; +} + +Object.values(fields).forEach((field) => field?.addEventListener('change', () => void save())); +document.querySelector('#shortcuts')?.addEventListener('click', () => void chrome.tabs.create({ url: 'chrome://extensions/shortcuts' })); +document.querySelector('#downloads')?.addEventListener('click', () => void chrome.tabs.create({ url: 'chrome://settings/downloads' })); +void load(); diff --git a/package.json b/package.json index d261bec..7e79b19 100644 --- a/package.json +++ b/package.json @@ -30,6 +30,10 @@ "lint:manifest": "node ./scripts/lint-manifest.mjs", "validate:store": "node ./scripts/validate-store-assets.mjs", "package:store": "pnpm build && pnpm lint:manifest && node ./scripts/package-chrome-store.mjs", + "entitlements:serve": "node ./services/entitlement-authority/server.mjs", + "entitlements:verify-key": "node ./scripts/verify-entitlement-key.mjs", + "site:build": "node ./scripts/build-site-config.mjs", + "package:helper:macos": "sh ./scripts/package-macos-helper.sh", "test:e2e": "playwright test --config tests/e2e/playwright.config.mjs", "test": "tsc -p tsconfig.json && node --test tests/panel/*.test.mjs && python3 -m unittest discover -s tests/server -p 'test_*.py'", "test:privacy": "node --test tests/panel/network-boundary.test.mjs && python3 -m unittest tests/server/test_privacy_proof.py", diff --git a/panel/panel.css b/panel/panel.css index b355e88..7b5f244 100644 --- a/panel/panel.css +++ b/panel/panel.css @@ -10,6 +10,7 @@ --accent: #0f172a; --accent-soft: rgba(15, 23, 42, 0.06); --success: #0f766e; + --context-tint: transparent; --radius: 22px; --panel-padding: 20px; --font: "SF Pro Display", "SF Pro Text", "Avenir Next", "Helvetica Neue", sans-serif; @@ -19,6 +20,7 @@ html, body { margin: 0; padding: 0; background: + radial-gradient(circle at 82% 8%, color-mix(in srgb, var(--context-tint) 8%, transparent), transparent 36%), radial-gradient(circle at top left, rgba(255,255,255,0.85), transparent 28%), linear-gradient(180deg, #fafafa 0%, #f2f4f7 100%); font-family: var(--font); @@ -26,6 +28,19 @@ html, body { min-height: 100vh; } +body[data-text-size="large"] { font-size: 108%; } +body[data-text-size="larger"] { font-size: 118%; } +body[data-high-contrast="true"] { --muted: #39433f; --panel-border: rgba(0, 0, 0, .24); } +body[data-control-side="left"] .memory-actions, +body[data-control-side="left"] .intent-actions, +body[data-control-side="left"] .secondary-grid { direction: rtl; } +body[data-control-side="left"] .memory-actions > *, +body[data-control-side="left"] .intent-actions > *, +body[data-control-side="left"] .secondary-grid > * { direction: ltr; } +body[data-reduce-motion="true"] *, +body[data-reduce-motion="true"] *::before, +body[data-reduce-motion="true"] *::after { animation: none !important; transition-duration: 0.01ms !important; } + .panel { width: 100%; box-sizing: border-box; @@ -100,13 +115,28 @@ html, body { color: var(--muted-strong); } +.tier:empty { display: none; } + +.panel-header { order: 1; } +.truth-strip { order: 2; } +.runtime-state { order: 3; } +.selection-shell { order: 4; } +.intent-shell { order: 5; } +.workspace { order: 6; } +.result-shell { order: 7; } +.exports { order: 8; } +.memory-shell { order: 9; } +.access-shell { order: 10; } +.status-footer { order: 11; } + .truth-strip { display: grid; - grid-template-columns: repeat(6, minmax(0, 1fr)); + grid-template-columns: repeat(3, minmax(0, 1fr)); gap: 8px; } .runtime-meta-overlay { + display: none; border: 1px solid rgba(15, 23, 42, 0.2); border-radius: 18px; background: rgba(15, 23, 42, 0.62); @@ -117,6 +147,57 @@ html, body { gap: 10px; } +.result-shell { position: relative; overflow: hidden; } + +.processing-field { + position: absolute; + right: 22px; + bottom: 20px; + width: 58px; + height: 42px; + opacity: 0; + pointer-events: none; + transition: opacity 240ms ease; +} + +.processing-field.is-active { opacity: 1; } + +.processing-field span { + position: absolute; + width: 28px; + height: 22px; + border-radius: 44% 56% 52% 48%; + background: color-mix(in srgb, var(--success) 22%, transparent); + filter: blur(7px); + opacity: .16; + transform: translate3d(0, 0, 0); + animation: field-drift-a 3.37s ease-in-out infinite alternate; +} + +.processing-field span:nth-child(2) { + right: 2px; + top: 8px; + animation-name: field-drift-b; + animation-duration: 3.91s; + animation-delay: -170ms; +} + +.processing-field span:nth-child(3) { + left: 16px; + bottom: 0; + animation-name: field-drift-c; + animation-duration: 4.13s; + animation-delay: -310ms; +} + +@keyframes field-drift-a { to { transform: translate3d(7px, 4px, 0); } } +@keyframes field-drift-b { to { transform: translate3d(-5px, 7px, 0); } } +@keyframes field-drift-c { to { transform: translate3d(4px, -6px, 0); } } + +@media (prefers-reduced-motion: reduce) { + .processing-field span { animation: none; } +} + .runtime-meta-header { display: flex; align-items: center; @@ -235,7 +316,7 @@ html, body { } .truth-chip { - padding: 12px; + padding: 10px 12px; display: grid; gap: 6px; } @@ -759,6 +840,23 @@ html, body { min-height: 20px; } +.installation-progress { + width: 100%; + height: 8px; + margin-top: 14px; + border-radius: 999px; + overflow: hidden; + background: rgba(15, 23, 42, 0.08); +} + +.installation-progress span { + display: block; + height: 100%; + border-radius: inherit; + background: linear-gradient(90deg, var(--success), rgba(15, 118, 110, 0.62)); + transition: width 480ms ease; +} + pre { margin: 0; white-space: pre-wrap; @@ -774,7 +872,6 @@ pre { } @media (max-width: 960px) { - .truth-strip, .runtime-grid, .profile-grid, .runtime-meta-grid { @@ -783,19 +880,13 @@ pre { } @media (max-width: 560px) { - .truth-strip, .runtime-grid, .profile-grid, .runtime-meta-grid, - .memory-actions, - .intent-actions, - .secondary-grid, - .advanced-grid { + .memory-actions { grid-template-columns: 1fr; } - .panel-header, - .title-row, .workspace-header, .result-header { align-items: flex-start; diff --git a/panel/panel.html b/panel/panel.html index 92d7af5..e4507e4 100644 --- a/panel/panel.html +++ b/panel/panel.html @@ -9,7 +9,7 @@
-
Privacy-first execution layer
+
Local text workspace
SelectPilot
@@ -18,35 +18,11 @@
-
-
-
-
Runtime meta
- Idle -
- Connecting… -
-

Waiting for deterministic local runtime events.

- -
-
Operation
-
Step
-
Trace
-
-
    -
    -
    Execution Detecting
    -
    - Model - Checking -
    Boundary Local-only @@ -56,14 +32,6 @@ Verifying Awaiting proof
    -
    - Profile - Pending -
    -
    - Health - Measuring -
    @@ -94,19 +62,23 @@

    Retention control

    -
    -
    +
    -
    Billing
    -

    Attach paid token

    +
    Access
    +

    Keep your flow available

    - +
    - - + +
    -

    No entitlement token attached yet.

    +
    + + +
    +

    No active access yet.

    Selection
    @@ -128,10 +129,6 @@

    Ready

    const selection = firstRunMode ? '' : 'Ship the update before Friday and inform the team.'; let memoryEnabled = false; const memoryLedger = []; - const mockOrders = { - 'SP-PAID': { paid: true, token: 'sp_paid_token_demo', confirmations: 12, confirmations_required: 10 }, - 'SP-PENDING': { paid: false, confirmations: 2, confirmations_required: 10 }, - }; let entitlement = null; function pushMemory(action) { @@ -241,7 +238,14 @@

    Ready

    return { markdown: 'ok', json: {} }; }, }, + tabs: { + async create() {}, + }, storage: { + sync: { + async get() { return {}; }, + async set() {}, + }, local: { _store: {}, async get(key) { @@ -262,15 +266,8 @@

    Ready

    globalThis.fetch = async (url, init = {}) => { const href = String(url); globalThis.__fetchLog.push({ href, method: init?.method || 'GET' }); - if (href.includes('127.0.0.1:8090/order/')) { - const orderId = href.split('/order/')[1] || ''; - const order = mockOrders[decodeURIComponent(orderId)]; - if (!order) { - return { ok: false, status: 404, json: async () => ({ error: 'not_found' }) }; - } - return { ok: true, status: 200, json: async () => ({ order_id: orderId, ...order }) }; - } const ok = (payload) => ({ ok: true, status: 200, json: async () => payload }); + if (href.includes('/license/trial')) return ok({ token: 'sp_trial_pro_demo' }); if (href.includes('/profiles')) { return ok({ profiles: [ diff --git a/tests/e2e/panel-user-flow.spec.mjs b/tests/e2e/panel-user-flow.spec.mjs index 0557ea7..253cf8a 100644 --- a/tests/e2e/panel-user-flow.spec.mjs +++ b/tests/e2e/panel-user-flow.spec.mjs @@ -23,16 +23,9 @@ test('sidepanel harness can run extract action and render result', async ({ page await page.click('#btn-memory-delete'); await expect(page.locator('#memory-status')).toContainText('0 retained events'); - await page.fill('#order-id', 'SP-PENDING'); - await page.click('#btn-sync-order'); - await expect(page.locator('#status')).toContainText('No payment detected yet'); - - await page.fill('#order-id', 'SP-PAID'); - await page.click('#btn-sync-order'); - await expect(page.locator('#entitlement-status')).toContainText('Token attached · tier pro'); - - await page.click('#btn-sync-order'); - await expect(page.locator('#status')).toContainText('Order already synced; entitlement refreshed'); + await page.click('#btn-start-trial'); + await expect(page.locator('#entitlement-status')).toContainText('pro access'); + await expect(page.locator('#status')).toContainText('Access ready'); }); test('unlocked first-run uses the dedicated example route and resolves once', async ({ page }) => { @@ -40,8 +33,8 @@ test('unlocked first-run uses the dedicated example route and resolves once', as await page.goto(`${pathToFileURL(harnessPath)}?firstRun=1`); await expect(page.locator('#btn-first-run-example')).toHaveCount(0); - await page.fill('#order-id', 'SP-PAID'); - await page.click('#btn-sync-order'); + await page.fill('#license-token', 'sp_paid_pro_demo'); + await page.click('#btn-attach-license'); await expect(page.locator('#btn-first-run-example')).toBeVisible(); await expect(page.locator('#selection-card')).toContainText('no page content used'); diff --git a/tests/panel/entitlement-authority.test.mjs b/tests/panel/entitlement-authority.test.mjs new file mode 100644 index 0000000..cd05a0a --- /dev/null +++ b/tests/panel/entitlement-authority.test.mjs @@ -0,0 +1,125 @@ +import assert from 'node:assert/strict'; +import { createHmac, generateKeyPairSync } from 'node:crypto'; +import { mkdtemp, writeFile } from 'node:fs/promises'; +import { tmpdir } from 'node:os'; +import path from 'node:path'; +import test from 'node:test'; +import { createAuthority, verifyPaddleSignature } from '../../services/entitlement-authority/server.mjs'; + +async function fixture(t) { + const root = await mkdtemp(path.join(tmpdir(), 'selectpilot-authority-')); + const statePath = path.join(root, 'state.json'); + const keyPath = path.join(root, 'private.pem'); + const { privateKey } = generateKeyPairSync('ed25519'); + await writeFile(keyPath, privateKey.export({ type: 'pkcs8', format: 'pem' }), { mode: 0o600 }); + const server = await createAuthority({ + statePath, + privateKeyFile: keyPath, + kid: 'test-2026-01', + webhookSecret: 'pdl_test_secret', + priceMap: { + __trial__: { tier: 'pro', features: ['extract', 'export'] }, + pri_plus: { tier: 'plus', features: ['extract', 'export'] }, + }, + }); + await new Promise((resolve) => server.listen(0, '127.0.0.1', resolve)); + t.after(() => new Promise((resolve) => server.close(resolve))); + return `http://127.0.0.1:${server.address().port}`; +} + +async function post(base, route, body, headers = {}) { + return fetch(`${base}${route}`, { + method: 'POST', + headers: { 'content-type': 'application/json', ...headers }, + body: typeof body === 'string' ? body : JSON.stringify(body), + }); +} + +test('trial is idempotent and produces a signed, expiring entitlement', async (t) => { + const base = await fixture(t); + const first = await post(base, '/v1/trials/start', { installation_id: 'installation-a' }); + assert.equal(first.status, 201); + const issued = await first.json(); + assert.match(issued.token, /^sp_trial_/); + + const repeated = await post(base, '/v1/trials/start', { installation_id: 'installation-a' }); + assert.equal((await repeated.json()).token, issued.token); + + const verified = await post(base, '/v1/entitlements/verify', { token: issued.token }); + assert.equal(verified.status, 200); + const signed = await verified.json(); + assert.equal(signed.entitlement.tier, 'pro'); + assert.equal(signed.alg, 'Ed25519'); + assert.equal(signed.kid, 'test-2026-01'); + assert.ok(signed.entitlement.expiresAt > Date.now()); + assert.match(signed.signature, /^[A-Za-z0-9+/]+=*$/); +}); + +test('Paddle webhook is authenticated and purchase claim redeems once', async (t) => { + const base = await fixture(t); + const event = JSON.stringify({ + event_id: 'evt_1', + event_type: 'transaction.completed', + data: { id: 'txn_1', custom_data: { claim_id: 'claim-secret' }, items: [{ price: { id: 'pri_plus' } }] }, + }); + const timestamp = Math.floor(Date.now() / 1000); + const digest = createHmac('sha256', 'pdl_test_secret').update(`${timestamp}:${event}`).digest('hex'); + const webhook = await post(base, '/v1/paddle/webhook', event, { 'paddle-signature': `ts=${timestamp};h1=${digest}` }); + assert.equal(webhook.status, 200); + + const claim = await post(base, '/v1/claims/redeem', { claim_id: 'claim-secret' }); + const ready = await claim.json(); + assert.equal(ready.status, 'ready'); + assert.match(ready.token, /^sp_live_/); + const repeated = await post(base, '/v1/claims/redeem', { claim_id: 'claim-secret' }); + assert.deepEqual(await repeated.json(), ready); + const acknowledged = await post(base, '/v1/claims/ack', { claim_id: 'claim-secret' }); + assert.deepEqual(await acknowledged.json(), { acknowledged: true }); + const afterAck = await post(base, '/v1/claims/redeem', { claim_id: 'claim-secret' }); + assert.deepEqual(await afterAck.json(), { status: 'acknowledged' }); +}); + +test('Paddle webhook rejects events without a stable subject', async (t) => { + const base = await fixture(t); + const event = JSON.stringify({ event_id: 'evt_missing_subject', event_type: 'transaction.completed', data: {} }); + const timestamp = Math.floor(Date.now() / 1000); + const digest = createHmac('sha256', 'pdl_test_secret').update(`${timestamp}:${event}`).digest('hex'); + const response = await post(base, '/v1/paddle/webhook', event, { 'paddle-signature': `ts=${timestamp};h1=${digest}` }); + assert.equal(response.status, 422); +}); + +test('subscription updates preserve the token and cancellation revokes it', async (t) => { + const base = await fixture(t); + const sendEvent = async (event) => { + const body = JSON.stringify(event); + const timestamp = Math.floor(Date.now() / 1000); + const digest = createHmac('sha256', 'pdl_test_secret').update(`${timestamp}:${body}`).digest('hex'); + return post(base, '/v1/paddle/webhook', body, { 'paddle-signature': `ts=${timestamp};h1=${digest}` }); + }; + const activated = await sendEvent({ + event_id: 'evt_activate', event_type: 'subscription.activated', + data: { id: 'sub_1', custom_data: { claim_id: 'claim-sub' }, items: [{ price: { id: 'pri_plus' } }] }, + }); + assert.equal(activated.status, 200); + const claimed = await (await post(base, '/v1/claims/redeem', { claim_id: 'claim-sub' })).json(); + + const updated = await sendEvent({ + event_id: 'evt_update', event_type: 'subscription.updated', + data: { id: 'sub_1', items: [{ price: { id: 'pri_plus' } }] }, + }); + assert.equal(updated.status, 200); + assert.equal((await post(base, '/v1/entitlements/verify', { token: claimed.token })).status, 200); + + const canceled = await sendEvent({ event_id: 'evt_cancel', event_type: 'subscription.canceled', data: { id: 'sub_1' } }); + assert.equal(canceled.status, 200); + assert.equal((await post(base, '/v1/entitlements/verify', { token: claimed.token })).status, 401); +}); + +test('Paddle signature rejects stale and forged events', () => { + const body = '{}'; + const now = 2_000_000_000; + const valid = createHmac('sha256', 'secret').update(`${now}:${body}`).digest('hex'); + assert.equal(verifyPaddleSignature(body, `ts=${now};h1=${valid}`, 'secret', now), true); + assert.equal(verifyPaddleSignature(body, `ts=${now};h1=${valid}`, 'wrong', now), false); + assert.equal(verifyPaddleSignature(body, `ts=${now - 301};h1=${valid}`, 'secret', now), false); +}); diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs index bd57949..70cc3d5 100644 --- a/tests/panel/store-package.test.mjs +++ b/tests/panel/store-package.test.mjs @@ -4,7 +4,8 @@ import assert from 'node:assert/strict'; import { execFileSync, spawnSync } from 'node:child_process'; import test from 'node:test'; -import { readFile } from 'node:fs/promises'; +import { mkdir, mkdtemp, readFile, writeFile } from 'node:fs/promises'; +import { tmpdir } from 'node:os'; import path from 'node:path'; import { fileURLToPath } from 'node:url'; import { collectRuntimeFiles, assertReleaseSafe } from '../../scripts/package-chrome-store.mjs'; @@ -26,6 +27,7 @@ test('runtime inventory excludes source, tests, reports, and transient files', a assert.ok(files.includes('content/content-script.bundle.js')); assert.ok(files.includes('assets/icon128.png')); assert.ok(files.includes('pricing/tier-feature-map.json')); + assert.ok(files.includes('pricing/entitlement-public-keys.json')); assert.ok(!files.some((file) => file.startsWith('billing/'))); assert.ok(!files.includes('pricing/paddle-products.json')); assert.ok(!files.some((file) => file.startsWith('assets/marketing/'))); @@ -33,15 +35,12 @@ test('runtime inventory excludes source, tests, reports, and transient files', a assert.ok(files.every((file) => !/(^|\/)(tests|reports|node_modules)(\/|$)/.test(file))); }); -test('store packaging fails closed while entitlement verification is unsigned', async () => { +test('store release accepts the pinned production entitlement keyring', async () => { const files = await collectRuntimeFiles(root); - await assert.rejects( - assertReleaseSafe(files, root), - /production entitlement signature verification is not configured/ - ); + await assert.doesNotReject(assertReleaseSafe(files, root)); }); -test('relative CLI paths execute validation and fail-closed packaging', () => { +test('relative CLI paths execute validation and deterministic packaging', () => { const validation = execFileSync(process.execPath, ['./scripts/validate-store-assets.mjs'], { cwd: root, encoding: 'utf8', @@ -52,6 +51,24 @@ test('relative CLI paths execute validation and fail-closed packaging', () => { cwd: root, encoding: 'utf8', }); - assert.equal(packaging.status, 1); - assert.match(packaging.stderr, /production entitlement signature verification is not configured/); + assert.equal(packaging.status, 0, packaging.stderr); + assert.match(packaging.stdout, /^selectpilot-1\.0\.0\.zip [0-9a-f]{64}\s*$/); +}); + +test('store release rejects remotely hosted executable code', async () => { + const releaseRoot = await mkdtemp(path.join(tmpdir(), 'selectpilot-remote-code-')); + await mkdir(path.join(releaseRoot, 'background'), { recursive: true }); + await writeFile( + path.join(releaseRoot, 'background/entitlement-service.js'), + 'const PUBLIC_KEY_HEX = "01";' + ); + await writeFile( + path.join(releaseRoot, 'remote.js'), + 'script.src = "https://cdn.example.com/runtime.js";' + ); + + await assert.rejects( + assertReleaseSafe(['remote.js'], releaseRoot), + /Remote hosted code/ + ); }); diff --git a/tests/server/test_entitlement_verification.py b/tests/server/test_entitlement_verification.py index d2f6d48..bed74be 100644 --- a/tests/server/test_entitlement_verification.py +++ b/tests/server/test_entitlement_verification.py @@ -48,19 +48,31 @@ def test_invalid_authority_token_remains_unauthorized(self) -> None: def test_verified_local_authority_response_is_returned(self) -> None: response = _Response( - b'{"entitlement":{"token":"opaque-token","tier":"pro","features":[],"issuedAt":1700000000000,' - b'"expiresAt":null},"signature":"c2lnbmF0dXJl","alg":"Ed25519","kid":"rotation-1"}' + b'{"entitlement":{"token":"opaque-token","tier":"pro","features":["extract"],' + b'"issuedAt":1700000000000,"expiresAt":null},"signature":"c2ln",' + b'"alg":"Ed25519","kid":"prod-1"}' ) with patch("nano_server.urlopen", return_value=response): result = license_verify({"token": "opaque-token"}) self.assertEqual(result["entitlement"]["tier"], "pro") - def test_non_loopback_verifier_configuration_is_rejected(self) -> None: - with patch.dict(os.environ, {"SELECTPILOT_BILLING_VERIFY_URL": "https://example.com/license/verify"}): + def test_insecure_remote_verifier_configuration_is_rejected(self) -> None: + with patch.dict(os.environ, {"SELECTPILOT_ENTITLEMENT_AUTHORITY_URL": "http://example.com"}): with self.assertRaises(ValidationError) as ctx: license_verify({"token": "opaque-token"}) self.assertEqual(ctx.exception.code, "invalid_entitlement_verifier") + def test_https_authority_is_allowed(self) -> None: + response = _Response( + b'{"entitlement":{"token":"opaque-token","tier":"plus","features":[],' + b'"issuedAt":1700000000000,"expiresAt":null},"signature":"c2ln",' + b'"alg":"Ed25519","kid":"prod-1"}' + ) + with patch.dict(os.environ, {"SELECTPILOT_ENTITLEMENT_AUTHORITY_URL": "https://license.selectpilot.app"}): + with patch("nano_server.urlopen", return_value=response) as request: + license_verify({"token": "opaque-token"}) + self.assertEqual(request.call_args.args[0].full_url, "https://license.selectpilot.app/v1/entitlements/verify") + def test_malformed_authority_response_is_rejected(self) -> None: response = _Response( b'{"entitlement":{"token":"opaque-token","tier":"pro","issuedAt":"not-a-timestamp"},' diff --git a/tests/server/test_installation_manager.py b/tests/server/test_installation_manager.py new file mode 100644 index 0000000..f4ac3f1 --- /dev/null +++ b/tests/server/test_installation_manager.py @@ -0,0 +1,39 @@ +"""Installation state and consent contract tests.""" +from __future__ import annotations + +import sys +import unittest +from pathlib import Path +from unittest.mock import patch + +ROOT = Path(__file__).resolve().parents[2] +SERVER_DIR = ROOT / "server" +if str(SERVER_DIR) not in sys.path: + sys.path.insert(0, str(SERVER_DIR)) + +from installation_manager import InstallationManager # noqa: E402 + + +class InstallationManagerTests(unittest.TestCase): + def test_consent_is_required(self) -> None: + manager = InstallationManager() + with self.assertRaisesRegex(ValueError, "installation_consent_required"): + manager.start(False) + + def test_non_macos_fails_before_work_starts(self) -> None: + manager = InstallationManager() + with patch("installation_manager.platform.system", return_value="Linux"): + with self.assertRaisesRegex(RuntimeError, "macos_required"): + manager.start(True) + self.assertEqual(manager.status()["state"], "idle") + + def test_repeated_start_is_single_flight(self) -> None: + manager = InstallationManager() + manager._state["state"] = "installing" + with patch("installation_manager.platform.system", return_value="Darwin"): + state = manager.start(True) + self.assertEqual(state["state"], "installing") + + +if __name__ == "__main__": + unittest.main() From 30f09e36f0937dcde40ffd460a5ba02a53a3c78a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:19:45 +0200 Subject: [PATCH 35/66] test(release): reach remote-code package guard Verified: full Hetzner release gate --- tests/panel/store-package.test.mjs | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs index 70cc3d5..8957951 100644 --- a/tests/panel/store-package.test.mjs +++ b/tests/panel/store-package.test.mjs @@ -58,10 +58,18 @@ test('relative CLI paths execute validation and deterministic packaging', () => test('store release rejects remotely hosted executable code', async () => { const releaseRoot = await mkdtemp(path.join(tmpdir(), 'selectpilot-remote-code-')); await mkdir(path.join(releaseRoot, 'background'), { recursive: true }); + await mkdir(path.join(releaseRoot, 'pricing'), { recursive: true }); await writeFile( path.join(releaseRoot, 'background/entitlement-service.js'), 'const PUBLIC_KEY_HEX = "01";' ); + await writeFile( + path.join(releaseRoot, 'pricing/entitlement-public-keys.json'), + JSON.stringify({ + schema_version: 1, + keys: [{ kid: 'test', alg: 'Ed25519', public_key_hex: '01'.repeat(32), status: 'active' }], + }) + ); await writeFile( path.join(releaseRoot, 'remote.js'), 'script.src = "https://cdn.example.com/runtime.js";' From 4669db6b113f170bbf9f0c5a78db6163e393f716 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:25:03 +0200 Subject: [PATCH 36/66] fix(ci): satisfy helper package shell lint Verified: shellcheck scripts/package-macos-helper.sh on Hetzner --- scripts/package-macos-helper.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/package-macos-helper.sh b/scripts/package-macos-helper.sh index ed9e713..c0a1916 100644 --- a/scripts/package-macos-helper.sh +++ b/scripts/package-macos-helper.sh @@ -1,7 +1,7 @@ #!/bin/sh set -eu -ROOT="$(CDPATH= cd -- "$(dirname "$0")/.." && pwd)" +ROOT="$(CDPATH='' cd -- "$(dirname "$0")/.." && pwd)" STAGE="$ROOT/dist/macos-helper/payload" OUTPUT="$ROOT/dist/macos-helper/SelectPilot-Installer-unsigned.pkg" SIGNED="$ROOT/dist/macos-helper/SelectPilot-Installer.pkg" From 1772eef8975f16f0a848e48583db2f9513bb85af Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:28:42 +0200 Subject: [PATCH 37/66] fix(ci): map release modules to product spec Verified: pnpm bench:frontier on Hetzner --- options/options.ts | 2 ++ scripts/build-site-config.mjs | 2 ++ scripts/verify-entitlement-key.mjs | 2 ++ server/installation_manager.py | 2 +- services/entitlement-authority/server.mjs | 3 +++ site/checkout.js | 2 ++ site/config.example.js | 2 ++ tests/panel/entitlement-authority.test.mjs | 2 ++ tests/server/test_installation_manager.py | 2 +- 9 files changed, 17 insertions(+), 2 deletions(-) diff --git a/options/options.ts b/options/options.ts index 456b6c1..5622be4 100644 --- a/options/options.ts +++ b/options/options.ts @@ -1,3 +1,5 @@ +// module_name: extension_settings +// spec_ref: "runtime_profiles" export {}; type Settings = { diff --git a/scripts/build-site-config.mjs b/scripts/build-site-config.mjs index 188e4d4..cf7b071 100644 --- a/scripts/build-site-config.mjs +++ b/scripts/build-site-config.mjs @@ -1,3 +1,5 @@ +// module_name: public_site_configuration +// spec_ref: "reporting" import { mkdir, readFile, writeFile } from 'node:fs/promises'; import path from 'node:path'; diff --git a/scripts/verify-entitlement-key.mjs b/scripts/verify-entitlement-key.mjs index a848f9c..328f8dd 100644 --- a/scripts/verify-entitlement-key.mjs +++ b/scripts/verify-entitlement-key.mjs @@ -1,3 +1,5 @@ +// module_name: entitlement_key_verification +// spec_ref: "validation_layer" import { createPrivateKey, createPublicKey, sign, verify } from 'node:crypto'; import { readFile } from 'node:fs/promises'; diff --git a/server/installation_manager.py b/server/installation_manager.py index a0f2810..2ae8163 100644 --- a/server/installation_manager.py +++ b/server/installation_manager.py @@ -1,4 +1,4 @@ -"""Calm, explicit local runtime provisioning for macOS.""" +"""module_name: installation_manager; spec_ref: "provisioning_layer".""" from __future__ import annotations import json diff --git a/services/entitlement-authority/server.mjs b/services/entitlement-authority/server.mjs index 4a0e888..494d193 100644 --- a/services/entitlement-authority/server.mjs +++ b/services/entitlement-authority/server.mjs @@ -1,3 +1,5 @@ +// module_name: entitlement_authority +// spec_ref: "validation_layer" import { createHash, createHmac, randomBytes, sign, timingSafeEqual } from 'node:crypto'; import { createServer } from 'node:http'; import { readFile, rename, writeFile } from 'node:fs/promises'; @@ -42,6 +44,7 @@ function parsePriceMap(value) { return parsed; } +// @spec_ref validation_layer export function verifyPaddleSignature(rawBody, header, secret, nowSeconds = Math.floor(Date.now() / 1000)) { const parts = Object.fromEntries(String(header || '').split(';').map((part) => part.split('=', 2))); const timestamp = Number(parts.ts); diff --git a/site/checkout.js b/site/checkout.js index ad3acec..7e81a95 100644 --- a/site/checkout.js +++ b/site/checkout.js @@ -1,3 +1,5 @@ +// module_name: public_checkout +// spec_ref: "validation_layer" (() => { const state = document.querySelector('#checkout-state'); const config = window.SELECTPILOT_CONFIG; diff --git a/site/config.example.js b/site/config.example.js index 69ca194..51cc57c 100644 --- a/site/config.example.js +++ b/site/config.example.js @@ -1,3 +1,5 @@ +// module_name: public_site_configuration_example +// spec_ref: "reporting" window.SELECTPILOT_CONFIG = { paddleEnvironment: "sandbox", paddleClientToken: "test_replace_me", diff --git a/tests/panel/entitlement-authority.test.mjs b/tests/panel/entitlement-authority.test.mjs index cd05a0a..50bb4a3 100644 --- a/tests/panel/entitlement-authority.test.mjs +++ b/tests/panel/entitlement-authority.test.mjs @@ -1,3 +1,5 @@ +// module_name: entitlement_authority_tests +// spec_ref: "testing_strategy.integration_tests" import assert from 'node:assert/strict'; import { createHmac, generateKeyPairSync } from 'node:crypto'; import { mkdtemp, writeFile } from 'node:fs/promises'; diff --git a/tests/server/test_installation_manager.py b/tests/server/test_installation_manager.py index f4ac3f1..0fe2354 100644 --- a/tests/server/test_installation_manager.py +++ b/tests/server/test_installation_manager.py @@ -1,4 +1,4 @@ -"""Installation state and consent contract tests.""" +"""module_name: installation_manager_tests; spec_ref: "testing_strategy.integration_tests".""" from __future__ import annotations import sys From 48bdbcc231976e19bce88c75b2a32e89d2e2d9d5 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:37:39 +0200 Subject: [PATCH 38/66] fix(privacy): restrict capture to selected text Remove undeclared whole-page and media collection from the shipped runtime and entitlement map. Package only the selected-text content bundle and fail tests if media capture returns. Verified: Hetzner pnpm lint, typecheck, build, test, test:privacy, package:store, test:e2e, bench:frontier --- README.md | 2 +- ZERO_LEAKAGE.md | 2 +- agent/agent-pipeline.js | 20 ++----- agent/agent-pipeline.ts | 18 ++----- agent/agent-types.ts | 4 +- background/background.js | 30 +++-------- background/background.ts | 30 +++-------- content/content-script.bundle.js | 62 ---------------------- content/content-script.js | 13 +---- content/content-script.ts | 22 +------- content/extract-audio.js | 12 ----- content/extract-audio.ts | 20 ------- content/extract-text.js | 12 ----- content/extract-text.ts | 13 ----- content/extract-video.js | 28 ---------- content/extract-video.ts | 36 ------------- docs/CHROME_WEB_STORE_SUBMISSION.md | 2 +- panel/panel.js | 7 +-- panel/panel.ts | 6 +-- pricing/tier-feature-map.json | 4 -- scripts/package-chrome-store.mjs | 1 + tests/panel/entitlement-signature.test.mjs | 2 +- tests/panel/store-package.test.mjs | 14 +++++ tests/server/test_entitlement_signer.py | 2 +- 24 files changed, 54 insertions(+), 308 deletions(-) delete mode 100644 content/extract-audio.js delete mode 100644 content/extract-audio.ts delete mode 100644 content/extract-video.js delete mode 100644 content/extract-video.ts diff --git a/README.md b/README.md index 239c915..1d6fc19 100644 --- a/README.md +++ b/README.md @@ -121,7 +121,7 @@ The repository configuration is authoritative for price and entitlement mapping: | --- | ---: | --- | | Essential | $1.99 | Local structured extraction, canonical metadata and manual copy/export | | Plus | $5.99 | Essential plus stateless summaries, batch clipping and connector-format exports | -| Pro | $14.99 | Plus plus multimodal processing and an explicit opt-in local knowledge layer | +| Pro | $14.99 | Plus plus advanced local reasoning and an explicit opt-in local knowledge layer | Essential and Plus do not retain a knowledge history. Pro stateful features must remain local, visible and user-controlled: retained data can be inspected, exported and deleted. A feature is available only when the signed entitlement and `pricing/tier-feature-map.json` both permit it; there is no silent downgrade or fallback. diff --git a/ZERO_LEAKAGE.md b/ZERO_LEAKAGE.md index 2c20479..d87b8e8 100644 --- a/ZERO_LEAKAGE.md +++ b/ZERO_LEAKAGE.md @@ -41,7 +41,7 @@ The runtime is also profile-based: the product prefers the smallest viable local ## Experimental surfaces -- `Transcribe` and `Vision OCR` remain prototype utilities and are not part of the privacy claim. +- Audio transcription and vision OCR are absent from the shipped runtime and entitlement map. - Billing code exists in the repo as a prototype but is not part of the selected-text local copilot flow. ## How to verify diff --git a/agent/agent-pipeline.js b/agent/agent-pipeline.js index 9d99b60..8d39ef7 100644 --- a/agent/agent-pipeline.js +++ b/agent/agent-pipeline.js @@ -1,15 +1,6 @@ // module_name: agent_agent-pipeline_ts // spec_ref: "execution_layer" import { runAgent } from './agent-client.js'; -function detectInput(context) { - if (context.media?.videoFrame) - return { kind: 'video', summary: 'video frame with OCR candidate' }; - if (context.media?.image) - return { kind: 'image', summary: 'image with OCR candidate' }; - if (context.media?.audio) - return { kind: 'audio', summary: 'audio snippet for transcription' }; - return { kind: 'text', summary: 'text or markdown selection' }; -} function classifyContent(markdown) { if (/```/m.test(markdown)) return 'technical_markdown'; @@ -40,23 +31,22 @@ function buildPrompt(normalized, context, contentClass, detected, userPrompt) { return `You are the local SelectPilot agent. Input type: ${detected}. Classification: ${contentClass}.\n${goal}\n${meta}\n\nContent:\n${normalized}`; } export async function runPipeline(input, context, userPrompt) { - const detected = detectInput(context); - const normalized = normalizeMarkdown(input || context.selection || context.pageText || context.markdown || '') || 'No content provided.'; + const normalized = normalizeMarkdown(input || context.selection || context.markdown || '') || 'No content provided.'; const contentClass = classifyContent(normalized); const chain = buildReasoningChain([ - `Detect input (${detected.summary})`, + 'Confirm selected text input', `Classify content (${contentClass})`, 'Normalize to canonical markdown', - 'Augment with metadata and optional multimodal references', + 'Augment with declared page metadata', 'Request structured response (JSON + Markdown) from local agent' ]); - const prompt = buildPrompt(normalized, context, contentClass, detected.kind, userPrompt); + const prompt = buildPrompt(normalized, context, contentClass, 'text', userPrompt); const agentOut = await runAgent(prompt, { ...context, markdown: normalized, metadata: { ...(context.metadata || {}), - detectedKind: detected.kind, + detectedKind: 'text', contentClass } }); diff --git a/agent/agent-pipeline.ts b/agent/agent-pipeline.ts index 4f87223..a71727b 100644 --- a/agent/agent-pipeline.ts +++ b/agent/agent-pipeline.ts @@ -3,13 +3,6 @@ import { runAgent } from './agent-client.js'; import type { AgentContext, AgentOutput } from './agent-types.js'; -function detectInput(context: AgentContext): { kind: 'text' | 'audio' | 'video' | 'image'; summary: string } { - if (context.media?.videoFrame) return { kind: 'video', summary: 'video frame with OCR candidate' }; - if (context.media?.image) return { kind: 'image', summary: 'image with OCR candidate' }; - if (context.media?.audio) return { kind: 'audio', summary: 'audio snippet for transcription' }; - return { kind: 'text', summary: 'text or markdown selection' }; -} - function classifyContent(markdown: string): string { if (/```/m.test(markdown)) return 'technical_markdown'; if (markdown.length > 1200) return 'longform'; @@ -45,24 +38,23 @@ function buildPrompt( } export async function runPipeline(input: string, context: AgentContext, userPrompt?: string): Promise { - const detected = detectInput(context); - const normalized = normalizeMarkdown(input || context.selection || context.pageText || context.markdown || '') || 'No content provided.'; + const normalized = normalizeMarkdown(input || context.selection || context.markdown || '') || 'No content provided.'; const contentClass = classifyContent(normalized); const chain = buildReasoningChain([ - `Detect input (${detected.summary})`, + 'Confirm selected text input', `Classify content (${contentClass})`, 'Normalize to canonical markdown', - 'Augment with metadata and optional multimodal references', + 'Augment with declared page metadata', 'Request structured response (JSON + Markdown) from local agent' ]); - const prompt = buildPrompt(normalized, context, contentClass, detected.kind, userPrompt); + const prompt = buildPrompt(normalized, context, contentClass, 'text', userPrompt); const agentOut = await runAgent(prompt, { ...context, markdown: normalized, metadata: { ...(context.metadata || {}), - detectedKind: detected.kind, + detectedKind: 'text', contentClass } }); diff --git a/agent/agent-types.ts b/agent/agent-types.ts index 892cdb3..dafa4ae 100644 --- a/agent/agent-types.ts +++ b/agent/agent-types.ts @@ -1,7 +1,7 @@ // module_name: agent_agent-types_ts // spec_ref: "execution_layer" export type DetectedInput = { - kind: 'text' | 'audio' | 'video' | 'image'; + kind: 'text'; content: string; metadata: Record; }; @@ -10,9 +10,7 @@ export type AgentContext = { url?: string; title?: string; selection?: string; - pageText?: string; markdown?: string; - media?: { audio?: string; videoFrame?: string; image?: string }; metadata?: Record; }; diff --git a/background/background.js b/background/background.js index 701f763..7a18f71 100644 --- a/background/background.js +++ b/background/background.js @@ -135,30 +135,16 @@ async function collectContext() { const tab = await getActiveTab(); if (!tab?.id) return {}; - const [selection, doc, audio, video] = await Promise.all([ - requestFromContent(tab.id, 'content:get_selection'), - requestFromContent(tab.id, 'content:get_document'), - requestFromContent(tab.id, 'content:get_audio'), - requestFromContent(tab.id, 'content:get_video') - ]); + const selection = await requestFromContent(tab.id, 'content:get_selection'); const selectionText = selection?.text?.text || ''; - const documentText = doc?.documentText?.text || ''; - const url = selection?.text?.url || doc?.documentText?.url || tab.url; - const title = selection?.text?.title || doc?.documentText?.title || tab.title; - const pageColor = selection?.text?.pageColor || doc?.documentText?.pageColor; + const url = selection?.text?.url || tab.url; + const title = selection?.text?.title || tab.title; + const pageColor = selection?.text?.pageColor; return { url: url || undefined, title: title || undefined, selection: selectionText || undefined, - pageText: documentText || undefined, - media: { - audio: audio?.audio?.audioUrl, - videoFrame: video?.video?.frame, - image: video?.video?.poster - }, metadata: { - audioDuration: audio?.audio?.duration, - videoDuration: video?.video?.duration, capturedAt: Date.now(), pageColor, } @@ -169,7 +155,7 @@ async function handleSummarize() { if (!allowed) throw new Error('Flow tier required for summarize'); const context = await collectContext(); - const text = context.selection || context.pageText || context.markdown || ''; + const text = context.selection || context.markdown || ''; const payload = { text, url: context.url, title: context.title, metadata: context.metadata }; const result = await summarize(payload); const sourceOrigin = context.url || 'local-context'; @@ -232,7 +218,7 @@ async function handleAgent(prompt) { if (!allowed) throw new Error('Flow tier required for ask/rewrite transforms'); const context = await collectContext(); - const content = context.selection || context.pageText || context.markdown || ''; + const content = context.selection || context.markdown || ''; const result = await runPipeline(content, context, prompt); const sourceOrigin = context.url || 'local-context'; await recordMemoryEvent({ @@ -254,7 +240,7 @@ async function handleSelectionPreview() { const context = await collectContext(); return { selection: context.selection || '', - pageText: context.pageText || '', + pageText: '', title: context.title || '', url: context.url || '', hasSelection: Boolean(context.selection && context.selection.trim()), @@ -269,7 +255,7 @@ async function handleIntentCompile(intent) { return compileIntent({ intent: trimmed, has_selection: Boolean(context.selection && context.selection.trim()), - has_page_text: Boolean(context.pageText && context.pageText.trim()), + has_page_text: false, }); } chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { diff --git a/background/background.ts b/background/background.ts index 1c4d6ed..d5c2154 100644 --- a/background/background.ts +++ b/background/background.ts @@ -160,32 +160,18 @@ async function requestFromContent(tabId: number, type: string): Promise { const tab = await getActiveTab(); if (!tab?.id) return {}; - const [selection, doc, audio, video] = await Promise.all([ - requestFromContent(tab.id, 'content:get_selection'), - requestFromContent(tab.id, 'content:get_document'), - requestFromContent(tab.id, 'content:get_audio'), - requestFromContent(tab.id, 'content:get_video') - ]); + const selection = await requestFromContent(tab.id, 'content:get_selection'); const selectionText = (selection as any)?.text?.text || ''; - const documentText = (doc as any)?.documentText?.text || ''; - const url = (selection as any)?.text?.url || (doc as any)?.documentText?.url || tab.url; - const title = (selection as any)?.text?.title || (doc as any)?.documentText?.title || tab.title; - const pageColor = (selection as any)?.text?.pageColor || (doc as any)?.documentText?.pageColor; + const url = (selection as any)?.text?.url || tab.url; + const title = (selection as any)?.text?.title || tab.title; + const pageColor = (selection as any)?.text?.pageColor; return { url: url || undefined, title: title || undefined, selection: selectionText || undefined, - pageText: documentText || undefined, - media: { - audio: (audio as any)?.audio?.audioUrl, - videoFrame: (video as any)?.video?.frame, - image: (video as any)?.video?.poster - }, metadata: { - audioDuration: (audio as any)?.audio?.duration, - videoDuration: (video as any)?.video?.duration, capturedAt: Date.now(), pageColor, } @@ -196,7 +182,7 @@ async function handleSummarize(): Promise { const { allowed } = await requireFeature('text_summarization'); if (!allowed) throw new Error('Flow tier required for summarize'); const context = await collectContext(); - const text = context.selection || context.pageText || context.markdown || ''; + const text = context.selection || context.markdown || ''; const payload = { text, url: context.url, title: context.title, metadata: context.metadata }; const result = await summarize(payload); const sourceOrigin = context.url || 'local-context'; @@ -260,7 +246,7 @@ async function handleAgent(prompt: string): Promise { const { allowed } = await requireFeature('basic_local_agent'); if (!allowed) throw new Error('Flow tier required for ask/rewrite transforms'); const context = await collectContext(); - const content = context.selection || context.pageText || context.markdown || ''; + const content = context.selection || context.markdown || ''; const result = await runPipeline(content, context, prompt); const sourceOrigin = context.url || 'local-context'; await recordMemoryEvent({ @@ -283,7 +269,7 @@ async function handleSelectionPreview(): Promise { const context = await collectContext(); return { selection: context.selection || '', - pageText: context.pageText || '', + pageText: '', title: context.title || '', url: context.url || '', hasSelection: Boolean(context.selection && context.selection.trim()), @@ -299,7 +285,7 @@ async function handleIntentCompile(intent: string): Promise { return compileIntent({ intent: trimmed, has_selection: Boolean(context.selection && context.selection.trim()), - has_page_text: Boolean(context.pageText && context.pageText.trim()), + has_page_text: false, }); } diff --git a/content/content-script.bundle.js b/content/content-script.bundle.js index 76dc8fc..6071515 100644 --- a/content/content-script.bundle.js +++ b/content/content-script.bundle.js @@ -20,59 +20,6 @@ pageColor: pageColor() }; } - function extractDocumentText() { - const article = document.querySelector("article"); - const target = article || document.body; - const text = clean(target.innerText || ""); - return { - text, - url: location.href, - title: document.title, - pageColor: pageColor() - }; - } - - // content/extract-audio.js - function extractAudio() { - const audio = document.querySelector("audio"); - if (!audio) - return null; - return { - audioUrl: audio.currentSrc || audio.src || void 0, - duration: Number.isFinite(audio.duration) ? audio.duration : void 0, - title: document.title, - pageUrl: location.href - }; - } - - // content/extract-video.js - async function captureFrame(video) { - try { - const canvas = document.createElement("canvas"); - canvas.width = video.videoWidth; - canvas.height = video.videoHeight; - const ctx = canvas.getContext("2d"); - if (!ctx) - return void 0; - ctx.drawImage(video, 0, 0, canvas.width, canvas.height); - return canvas.toDataURL("image/png"); - } catch (e) { - console.error("frame capture failed", e); - return void 0; - } - } - async function extractVideo() { - const video = document.querySelector("video"); - if (!video) - return null; - return { - poster: video.poster || void 0, - currentTime: video.currentTime, - duration: Number.isFinite(video.duration) ? video.duration : void 0, - frame: await captureFrame(video), - pageUrl: location.href - }; - } // utils/logger.js var ENABLE_LOG = true; @@ -89,15 +36,6 @@ if (msg.type === "content:get_selection") { response.text = extractSelection(); } - if (msg.type === "content:get_document") { - response.documentText = extractDocumentText(); - } - if (msg.type === "content:get_audio") { - response.audio = extractAudio(); - } - if (msg.type === "content:get_video") { - response.video = await extractVideo(); - } sendResponse(response); })(); return true; diff --git a/content/content-script.js b/content/content-script.js index 2b12dc9..d2770a9 100644 --- a/content/content-script.js +++ b/content/content-script.js @@ -1,8 +1,6 @@ // module_name: content_content-script_ts // spec_ref: "execution_layer" -import { extractSelection, extractDocumentText } from './extract-text.js'; -import { extractAudio } from './extract-audio.js'; -import { extractVideo } from './extract-video.js'; +import { extractSelection } from './extract-text.js'; import { log } from '../utils/logger.js'; chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { (async () => { @@ -10,15 +8,6 @@ chrome.runtime.onMessage.addListener((msg, _sender, sendResponse) => { if (msg.type === 'content:get_selection') { response.text = extractSelection(); } - if (msg.type === 'content:get_document') { - response.documentText = extractDocumentText(); - } - if (msg.type === 'content:get_audio') { - response.audio = extractAudio(); - } - if (msg.type === 'content:get_video') { - response.video = await extractVideo(); - } sendResponse(response); })(); return true; diff --git a/content/content-script.ts b/content/content-script.ts index 99428aa..190f20a 100644 --- a/content/content-script.ts +++ b/content/content-script.ts @@ -1,21 +1,12 @@ // module_name: content_content-script_ts // spec_ref: "execution_layer" -import { extractSelection, extractDocumentText } from './extract-text.js'; -import { extractAudio } from './extract-audio.js'; -import { extractVideo } from './extract-video.js'; +import { extractSelection } from './extract-text.js'; import { log } from '../utils/logger.js'; -type Message = - | { type: 'content:get_selection' } - | { type: 'content:get_document' } - | { type: 'content:get_audio' } - | { type: 'content:get_video' }; +type Message = { type: 'content:get_selection' }; type Response = { text?: ReturnType; - documentText?: ReturnType; - audio?: Awaited>; - video?: Awaited>; }; chrome.runtime.onMessage.addListener((msg: Message, _sender, sendResponse) => { @@ -24,15 +15,6 @@ chrome.runtime.onMessage.addListener((msg: Message, _sender, sendResponse) => { if (msg.type === 'content:get_selection') { response.text = extractSelection(); } - if (msg.type === 'content:get_document') { - response.documentText = extractDocumentText(); - } - if (msg.type === 'content:get_audio') { - response.audio = extractAudio(); - } - if (msg.type === 'content:get_video') { - response.video = await extractVideo(); - } sendResponse(response); })(); return true; diff --git a/content/extract-audio.js b/content/extract-audio.js deleted file mode 100644 index d2149f4..0000000 --- a/content/extract-audio.js +++ /dev/null @@ -1,12 +0,0 @@ -// @spec_ref execution_layer -export function extractAudio() { - const audio = document.querySelector('audio'); - if (!audio) - return null; - return { - audioUrl: audio.currentSrc || audio.src || undefined, - duration: Number.isFinite(audio.duration) ? audio.duration : undefined, - title: document.title, - pageUrl: location.href - }; -} diff --git a/content/extract-audio.ts b/content/extract-audio.ts deleted file mode 100644 index 44d08c9..0000000 --- a/content/extract-audio.ts +++ /dev/null @@ -1,20 +0,0 @@ -// module_name: content_extract-audio_ts -// spec_ref: "execution_layer" -export type AudioExtraction = { - audioUrl?: string; - duration?: number; - title?: string; - pageUrl: string; -}; - -// @spec_ref execution_layer -export function extractAudio(): AudioExtraction | null { - const audio = document.querySelector('audio'); - if (!audio) return null; - return { - audioUrl: audio.currentSrc || audio.src || undefined, - duration: Number.isFinite(audio.duration) ? audio.duration : undefined, - title: document.title, - pageUrl: location.href - }; -} diff --git a/content/extract-text.js b/content/extract-text.js index 1a715c3..8850c1c 100644 --- a/content/extract-text.js +++ b/content/extract-text.js @@ -18,15 +18,3 @@ export function extractSelection() { pageColor: pageColor(), }; } -// @spec_ref execution_layer -export function extractDocumentText() { - const article = document.querySelector('article'); - const target = article || document.body; - const text = clean(target.innerText || ''); - return { - text, - url: location.href, - title: document.title, - pageColor: pageColor(), - }; -} diff --git a/content/extract-text.ts b/content/extract-text.ts index 1b242ea..34c0f50 100644 --- a/content/extract-text.ts +++ b/content/extract-text.ts @@ -28,16 +28,3 @@ export function extractSelection(): TextExtraction { pageColor: pageColor(), }; } - -// @spec_ref execution_layer -export function extractDocumentText(): TextExtraction { - const article = document.querySelector('article'); - const target = article || document.body; - const text = clean(target.innerText || ''); - return { - text, - url: location.href, - title: document.title, - pageColor: pageColor(), - }; -} diff --git a/content/extract-video.js b/content/extract-video.js deleted file mode 100644 index aaeb73e..0000000 --- a/content/extract-video.js +++ /dev/null @@ -1,28 +0,0 @@ -async function captureFrame(video) { - try { - const canvas = document.createElement('canvas'); - canvas.width = video.videoWidth; - canvas.height = video.videoHeight; - const ctx = canvas.getContext('2d'); - if (!ctx) - return undefined; - ctx.drawImage(video, 0, 0, canvas.width, canvas.height); - return canvas.toDataURL('image/png'); - } - catch (e) { - console.error('frame capture failed', e); - return undefined; - } -} -export async function extractVideo() { - const video = document.querySelector('video'); - if (!video) - return null; - return { - poster: video.poster || undefined, - currentTime: video.currentTime, - duration: Number.isFinite(video.duration) ? video.duration : undefined, - frame: await captureFrame(video), - pageUrl: location.href - }; -} diff --git a/content/extract-video.ts b/content/extract-video.ts deleted file mode 100644 index 04670fe..0000000 --- a/content/extract-video.ts +++ /dev/null @@ -1,36 +0,0 @@ -// module_name: content_extract-video_ts -// spec_ref: "execution_layer" -export type VideoExtraction = { - poster?: string; - currentTime?: number; - duration?: number; - frame?: string; - pageUrl: string; -}; - -async function captureFrame(video: HTMLVideoElement): Promise { - try { - const canvas = document.createElement('canvas'); - canvas.width = video.videoWidth; - canvas.height = video.videoHeight; - const ctx = canvas.getContext('2d'); - if (!ctx) return undefined; - ctx.drawImage(video, 0, 0, canvas.width, canvas.height); - return canvas.toDataURL('image/png'); - } catch (e) { - console.error('frame capture failed', e); - return undefined; - } -} - -export async function extractVideo(): Promise { - const video = document.querySelector('video'); - if (!video) return null; - return { - poster: (video as HTMLVideoElement).poster || undefined, - currentTime: video.currentTime, - duration: Number.isFinite(video.duration) ? video.duration : undefined, - frame: await captureFrame(video), - pageUrl: location.href - }; -} diff --git a/docs/CHROME_WEB_STORE_SUBMISSION.md b/docs/CHROME_WEB_STORE_SUBMISSION.md index 2fe0229..f525ff7 100644 --- a/docs/CHROME_WEB_STORE_SUBMISSION.md +++ b/docs/CHROME_WEB_STORE_SUBMISSION.md @@ -51,7 +51,7 @@ These answers must be rechecked against the exact submitted SHA and the hosted p ### Data handled for core functionality -- Website content: selected text and active-page context chosen by the user +- Website content: text explicitly selected by the user, plus the active page title and URL - User activity: the user's explicit SelectPilot actions and locally stored feature state - Authentication information: local entitlement token and signed entitlement metadata diff --git a/panel/panel.js b/panel/panel.js index b7638ff..07da0c5 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -987,18 +987,15 @@ function renderSelectionState() { else if (selectionPreview.hasSelection) { copy.textContent = shorten(selectionPreview.selection, 260); } - else if (selectionPreview.pageText) { - copy.textContent = 'Extract JSON requires highlighted text. Summarize and Ask can still fall back to page context.'; - } else { copy.textContent = 'Highlight text on the current page to activate the primary action.'; } const meta = document.createElement('p'); meta.className = 'selection-copy'; - const charCount = selectionPreview.hasSelection ? selectionPreview.selection.length : selectionPreview.pageText.length; + const charCount = selectionPreview.selection.length; meta.textContent = showFirstRunExample ? 'Local example · Action Brief · no page content used' - : `${selectionPreview.hasSelection ? 'Selection' : 'Page context'} · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; + : `Selection · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; selectionCardEl.append(header, title, copy, meta); if (showFirstRunExample) { const action = document.createElement('button'); diff --git a/panel/panel.ts b/panel/panel.ts index c292a7c..d5b0c79 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -1138,18 +1138,16 @@ function renderSelectionState() { copy.textContent = FIRST_RUN_EXAMPLE.text; } else if (selectionPreview.hasSelection) { copy.textContent = shorten(selectionPreview.selection, 260); - } else if (selectionPreview.pageText) { - copy.textContent = 'Extract JSON requires highlighted text. Summarize and Ask can still fall back to page context.'; } else { copy.textContent = 'Highlight text on the current page to activate the primary action.'; } const meta = document.createElement('p'); meta.className = 'selection-copy'; - const charCount = selectionPreview.hasSelection ? selectionPreview.selection.length : selectionPreview.pageText.length; + const charCount = selectionPreview.selection.length; meta.textContent = showFirstRunExample ? 'Local example · Action Brief · no page content used' - : `${selectionPreview.hasSelection ? 'Selection' : 'Page context'} · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; + : `Selection · ${charCount} chars${selectionPreview.url ? ` · ${selectionPreview.url}` : ''}`; selectionCardEl.append(header, title, copy, meta); if (showFirstRunExample) { diff --git a/pricing/tier-feature-map.json b/pricing/tier-feature-map.json index 3a92cb2..7ee2970 100644 --- a/pricing/tier-feature-map.json +++ b/pricing/tier-feature-map.json @@ -21,10 +21,6 @@ "structured_summaries" ], "pro": [ - "audio_transcription", - "video_frame_ocr", - "image_ocr", - "multimodal_clipper", "local_embeddings", "advanced_agent_reasoning", "project_memory", diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs index 626e210..e089a6c 100644 --- a/scripts/package-chrome-store.mjs +++ b/scripts/package-chrome-store.mjs @@ -51,6 +51,7 @@ export async function collectRuntimeFiles(root = projectRoot) { const absoluteRoot = path.join(root, runtimeRoot); for (const nested of await walk(absoluteRoot)) { const relative = path.posix.join(runtimeRoot, nested.split(path.sep).join(path.posix.sep)); + if (runtimeRoot === 'content' && relative !== 'content/content-script.bundle.js') continue; if (allowedExtensions.has(path.extname(relative)) && !forbiddenPath.test(relative)) files.push(relative); } } diff --git a/tests/panel/entitlement-signature.test.mjs b/tests/panel/entitlement-signature.test.mjs index 5dbc532..7e6a8d3 100644 --- a/tests/panel/entitlement-signature.test.mjs +++ b/tests/panel/entitlement-signature.test.mjs @@ -15,7 +15,7 @@ test('only a matching Ed25519 key ID and exact payload verify', async () => { const entitlement = { token: 'opaque-token', tier: 'pro', - features: ['image_ocr'], + features: ['project_memory'], issuedAt: 1_700_000_000_000, expiresAt: 4_000_000_000_000, }; diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs index 8957951..4e461a1 100644 --- a/tests/panel/store-package.test.mjs +++ b/tests/panel/store-package.test.mjs @@ -25,12 +25,14 @@ test('runtime inventory excludes source, tests, reports, and transient files', a assert.ok(files.includes('manifest.json')); assert.ok(files.includes('background/background.js')); assert.ok(files.includes('content/content-script.bundle.js')); + assert.ok(!files.some((file) => file.startsWith('content/') && file !== 'content/content-script.bundle.js')); assert.ok(files.includes('assets/icon128.png')); assert.ok(files.includes('pricing/tier-feature-map.json')); assert.ok(files.includes('pricing/entitlement-public-keys.json')); assert.ok(!files.some((file) => file.startsWith('billing/'))); assert.ok(!files.includes('pricing/paddle-products.json')); assert.ok(!files.some((file) => file.startsWith('assets/marketing/'))); + assert.ok(!files.some((file) => /(?:audio|video|image).*(?:extract|ocr|transcri)/i.test(file))); assert.ok(files.every((file) => !file.endsWith('.ts'))); assert.ok(files.every((file) => !/(^|\/)(tests|reports|node_modules)(\/|$)/.test(file))); }); @@ -40,6 +42,18 @@ test('store release accepts the pinned production entitlement keyring', async () await assert.doesNotReject(assertReleaseSafe(files, root)); }); +test('store runtime captures selected text only', async () => { + const bundle = await readFile(path.join(root, 'content/content-script.bundle.js'), 'utf8'); + const background = await readFile(path.join(root, 'background/background.js'), 'utf8'); + assert.match(bundle, /content:get_selection/); + assert.doesNotMatch(bundle, /content:get_(?:document|audio|video)/); + assert.doesNotMatch(bundle, /querySelector\(["'](?:audio|video)["']\)|drawImage\(/); + assert.doesNotMatch(background, /content:get_(?:document|audio|video)/); + + const tiers = JSON.parse(await readFile(path.join(root, 'pricing/tier-feature-map.json'), 'utf8')); + assert.ok(Object.values(tiers).flat().every((feature) => !/(?:audio|video|image|multimodal)/i.test(feature))); +}); + test('relative CLI paths execute validation and deterministic packaging', () => { const validation = execFileSync(process.execPath, ['./scripts/validate-store-assets.mjs'], { cwd: root, diff --git a/tests/server/test_entitlement_signer.py b/tests/server/test_entitlement_signer.py index b6a0767..777ca22 100644 --- a/tests/server/test_entitlement_signer.py +++ b/tests/server/test_entitlement_signer.py @@ -20,7 +20,7 @@ class EntitlementSignerTests(unittest.TestCase): entitlement = { "token": "opaque-token", "tier": "pro", - "features": ["image_ocr"], + "features": ["project_memory"], "issuedAt": 1_700_000_000_000, "expiresAt": 1_700_086_400_000, } From 547782b05c3dedffd3ff5eb7e81edfae152555f0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 17:59:34 +0200 Subject: [PATCH 39/66] fix(runtime): bind deterministic sampling and prewarm Use zero-temperature seeded generation across structured routes and prepare the selected model before installation resolves. Verified: Hetzner pnpm lint, typecheck, build, test, test:privacy, package:store, test:e2e, stress runner, bench:frontier, and shellcheck. --- README.md | 2 +- docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md | 2 ++ docs/model-runtime-decisions.md | 3 ++ .../com.selectpilot.bridge.plist.template | 1 + installer/macos/scripts/postinstall | 1 + launchd/com.chromeai.nano.plist | 1 + scripts/bootstrap-macos-local.sh | 35 +++++++++++++++++++ scripts/install-macos-local.sh | 9 +++-- server/installation_manager.py | 29 ++++++++++++++- server/ollama_client.py | 28 ++++++++++++--- tests/server/test_installation_manager.py | 33 ++++++++++++++++- tests/server/test_ollama_client.py | 3 ++ 12 files changed, 136 insertions(+), 11 deletions(-) diff --git a/README.md b/README.md index 1d6fc19..179522d 100644 --- a/README.md +++ b/README.md @@ -47,7 +47,7 @@ SelectPilot is built to keep the core workflow local-first and inspectable. pnpm setup:local ``` -This installs dependencies, builds the extension, installs/starts Ollama when needed, selects a hardware-safe local profile, pulls the exact models, installs the LaunchAgent, and refuses to finish unless the bridge and both configured models are healthy. +This installs dependencies, builds the extension, installs/starts Ollama when needed, selects a hardware-safe local profile, pulls and prewarms the exact generation model, installs the LaunchAgent, and refuses to finish unless the bridge and both configured models are healthy. Structured generation uses a fixed zero-temperature sampling contract and seed so repeated requests follow the same decision path. 2. Open `chrome://extensions`, enable **Developer mode**, choose **Load unpacked**, and select this repository folder. 3. Highlight text on a page → open SelectPilot → click **Extract JSON**. diff --git a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md index 3ee75e8..b9b2f8a 100644 --- a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md +++ b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md @@ -59,6 +59,8 @@ The separately operated local billing authority installs its pinned runtime with - [ ] The extension cannot read the active tab URL before either user gesture. - [ ] Missing Ollama produces one calm, actionable installation path. - [ ] Consent-gated model provisioning succeeds on supported hardware. +- [ ] Installation prewarms the selected generation model before reporting Ready. +- [ ] Two identical warm requests produce identical canonical JSON with the fixed production seed. - [ ] First selected-text extraction reaches a validated result. - [ ] JSON, Markdown, and plain-text exports open correctly. - [ ] Failure and recovery paths are verified without silent fallback. diff --git a/docs/model-runtime-decisions.md b/docs/model-runtime-decisions.md index e283e31..59abd14 100644 --- a/docs/model-runtime-decisions.md +++ b/docs/model-runtime-decisions.md @@ -6,9 +6,12 @@ - Balanced / Apple Silicon with 16 GB or more: `gemma4:e4b-it-qat`, `num_ctx=32768`. - Advanced: manual opt-in only, `qwen2.5:7b`, `num_ctx=32768`. - Embeddings: `nomic-embed-text-v2-moe:latest`. +- Structured generation: `temperature=0` and `seed=42` for every production profile. Explicit environment overrides remain authoritative and must fail visibly when invalid or unavailable. SelectPilot never substitutes a different installed model, downloads a model without consent, or routes generation to a cloud model. +Installation prewarms the selected generation model with the same context window and deterministic sampling options used at runtime. A failed prewarm is an explicit installation failure; SelectPilot does not report the runtime as ready and does not silently choose another model. `CHROMEAI_OLLAMA_SEED` may override the seed for controlled verification, but invalid or negative values fail closed. + ## Deferred decisions - Keep `nomic-embed-text-v2-moe:latest` until measured retrieval quality makes embeddings a product bottleneck (SOD-397/SOD-401). diff --git a/installer/macos/com.selectpilot.bridge.plist.template b/installer/macos/com.selectpilot.bridge.plist.template index 4876fd2..4d8f375 100644 --- a/installer/macos/com.selectpilot.bridge.plist.template +++ b/installer/macos/com.selectpilot.bridge.plist.template @@ -22,6 +22,7 @@ CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ CHROMEAI_OLLAMA_EMBED_MODEL__OLLAMA_EMBED_MODEL__ CHROMEAI_OLLAMA_NUM_CTX__OLLAMA_NUM_CTX__ + CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_AUTO_PULL_MODELS0 SELECTPILOT_ENTITLEMENT_AUTHORITY_URL__ENTITLEMENT_AUTHORITY_URL__ diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall index 7eeba2e..1408c7b 100644 --- a/installer/macos/scripts/postinstall +++ b/installer/macos/scripts/postinstall @@ -37,6 +37,7 @@ sed \ -e "s|__OLLAMA_MODEL__|$OLLAMA_MODEL|g" \ -e "s|__OLLAMA_EMBED_MODEL__|$OLLAMA_EMBED_MODEL|g" \ -e "s|__OLLAMA_NUM_CTX__|$OLLAMA_NUM_CTX|g" \ + -e "s|__OLLAMA_SEED__|42|g" \ -e "s|__ENTITLEMENT_AUTHORITY_URL__|https://license.selectpilot.app|g" \ "$SOURCE_ROOT/launchd/com.selectpilot.bridge.plist.template" > "$AGENT_PATH" chown "$CONSOLE_USER" "$AGENT_PATH" diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index 13f4506..45c1b71 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -26,6 +26,7 @@ CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ CHROMEAI_OLLAMA_EMBED_MODEL__OLLAMA_EMBED_MODEL__ CHROMEAI_OLLAMA_NUM_CTX__OLLAMA_NUM_CTX__ + CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_AUTO_PULL_MODELS0 diff --git a/scripts/bootstrap-macos-local.sh b/scripts/bootstrap-macos-local.sh index 9969ae5..7882e39 100755 --- a/scripts/bootstrap-macos-local.sh +++ b/scripts/bootstrap-macos-local.sh @@ -11,6 +11,7 @@ BRIDGE_HEALTH_URL="http://127.0.0.1:8083/health" STATUS_OLLAMA_INSTALL="pending" STATUS_OLLAMA_RUNNING="pending" STATUS_MODEL_PULL="pending" +STATUS_MODEL_WARMUP="pending" STATUS_LAUNCHAGENT="pending" STATUS_BRIDGE_HEALTH="pending" @@ -138,9 +139,42 @@ else STATUS_MODEL_PULL="skipped" fi +echo "Preparing generation model: $GENERATION_MODEL" +python3 - "$GENERATION_MODEL" "$NUM_CTX" <<'PY' +import json +import sys +from urllib.request import Request, urlopen + +model = sys.argv[1] +num_ctx = int(sys.argv[2]) +request = Request( + "http://127.0.0.1:11434/api/generate", + data=json.dumps({ + "model": model, + "prompt": "Return only: ready", + "stream": False, + "keep_alive": "10m", + "options": { + "temperature": 0, + "seed": 42, + "num_ctx": num_ctx, + "num_predict": 8, + }, + }).encode("utf-8"), + headers={"Content-Type": "application/json"}, + method="POST", +) +with urlopen(request, timeout=600) as response: + result = json.load(response) +if result.get("done") is not True: + raise SystemExit("Generation model did not finish preparing.") +PY +STATUS_MODEL_WARMUP="ok" + CHROMEAI_OLLAMA_MODEL="$GENERATION_MODEL" \ CHROMEAI_OLLAMA_EMBED_MODEL="$EMBEDDING_MODEL" \ CHROMEAI_OLLAMA_NUM_CTX="$NUM_CTX" \ +CHROMEAI_OLLAMA_SEED=42 \ "$ROOT/scripts/install-macos-local.sh" STATUS_LAUNCHAGENT="ok" @@ -187,6 +221,7 @@ Next recommended command: Bootstrap report: Ollama install: $STATUS_OLLAMA_INSTALL Ollama running: $STATUS_OLLAMA_RUNNING + Model prepared: $STATUS_MODEL_WARMUP Model pull: $STATUS_MODEL_PULL LaunchAgent: $STATUS_LAUNCHAGENT Bridge health: $STATUS_BRIDGE_HEALTH diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index c10c2f1..cd31309 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -9,6 +9,7 @@ OLLAMA_BASE_URL="${CHROMEAI_OLLAMA_BASE_URL:-http://127.0.0.1:11434}" OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-gemma4:e2b-it-qat}" OLLAMA_EMBED_MODEL="${CHROMEAI_OLLAMA_EMBED_MODEL:-nomic-embed-text-v2-moe:latest}" OLLAMA_NUM_CTX="${CHROMEAI_OLLAMA_NUM_CTX:-16384}" +OLLAMA_SEED="${CHROMEAI_OLLAMA_SEED:-42}" RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run}" LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}" @@ -24,6 +25,7 @@ sed \ -e "s|__OLLAMA_MODEL__|$OLLAMA_MODEL|g" \ -e "s|__OLLAMA_EMBED_MODEL__|$OLLAMA_EMBED_MODEL|g" \ -e "s|__OLLAMA_NUM_CTX__|$OLLAMA_NUM_CTX|g" \ + -e "s|__OLLAMA_SEED__|$OLLAMA_SEED|g" \ "$TEMPLATE" > "$DEST" launchctl unload "$DEST" 2>/dev/null || true @@ -40,7 +42,8 @@ Next steps: 3. Current Ollama model: $OLLAMA_MODEL 4. Local bridge URL: http://127.0.0.1:8083 5. Ollama context window: $OLLAMA_NUM_CTX - 6. Run dir: $RUN_DIR - 7. Log dir: $LOG_DIR - 8. Run 'pnpm benchmark:local' to validate latency on this machine. + 6. Deterministic seed: $OLLAMA_SEED + 7. Run dir: $RUN_DIR + 8. Log dir: $LOG_DIR + 9. Run 'pnpm benchmark:local' to validate latency on this machine. EOF diff --git a/server/installation_manager.py b/server/installation_manager.py index 2ae8163..f347a46 100644 --- a/server/installation_manager.py +++ b/server/installation_manager.py @@ -126,6 +126,32 @@ def _pull_model(self, model: str, start: int, end: int) -> None: if event.get("error"): raise RuntimeError("model_download_failed") + def _warm_model(self, model: str, num_ctx: int) -> None: + seed = int(os.environ.get("CHROMEAI_OLLAMA_SEED", "42")) + if seed < 0: + raise RuntimeError("model_warmup_failed") + request = Request( + "http://127.0.0.1:11434/api/generate", + data=json.dumps({ + "model": model, + "prompt": "Return only: ready", + "stream": False, + "keep_alive": "10m", + "options": { + "temperature": 0, + "seed": seed, + "num_ctx": num_ctx, + "num_predict": 8, + }, + }).encode("utf-8"), + headers={"Content-Type": "application/json"}, + method="POST", + ) + with urlopen(request, timeout=600) as response: + result = json.load(response) + if result.get("done") is not True: + raise RuntimeError("model_warmup_failed") + def _run(self) -> None: try: recommendation = recommend_runtime_profile() @@ -139,11 +165,12 @@ def _run(self) -> None: self._pull_model(profile.generation_model, 48, 82) self._pull_model(profile.embedding_model, 82, 94) self._update(label="Final checks", progress=96) + self._warm_model(profile.generation_model, profile.num_ctx) time.sleep(0.2) self._update(state="ready", label="Go", progress=100) except Exception as error: code = str(error) if str(error) in { - "ollama_start_timeout", "model_download_failed", "macos_required" + "ollama_start_timeout", "model_download_failed", "model_warmup_failed", "macos_required" } else "installation_failed" self._update( state="action_required", diff --git a/server/ollama_client.py b/server/ollama_client.py index 489c151..98cc076 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -98,6 +98,7 @@ class OllamaConfig: embed_model: str timeout_seconds: float num_ctx: int + seed: int class OllamaError(RuntimeError): @@ -114,6 +115,16 @@ def _positive_int(value: Any, name: str) -> int: return parsed +def _nonnegative_int(value: Any, name: str) -> int: + try: + parsed = int(value) + except (TypeError, ValueError) as exc: + raise OllamaError(f"{name} must be a non-negative integer") from exc + if parsed < 0: + raise OllamaError(f"{name} must be a non-negative integer") + return parsed + + class OllamaClient: def __init__(self, config: OllamaConfig | None = None): if config is None: @@ -137,7 +148,8 @@ def __init__(self, config: OllamaConfig | None = None): embed_model=os.environ.get("CHROMEAI_OLLAMA_EMBED_MODEL", default_embed_model), timeout_seconds=float(os.environ.get("CHROMEAI_OLLAMA_TIMEOUT_SECONDS", "30")), num_ctx=_positive_int(os.environ.get("CHROMEAI_OLLAMA_NUM_CTX", default_num_ctx), "CHROMEAI_OLLAMA_NUM_CTX"), - ) + seed=_nonnegative_int(os.environ.get("CHROMEAI_OLLAMA_SEED", "42"), "CHROMEAI_OLLAMA_SEED"), + ) self.config = config def _model_available_locally(self, requested: str, local_models: list[str]) -> bool: @@ -233,7 +245,11 @@ def _resolve_model(self, requested: str, preferences: list[str], models: list[st return requested def _generation_options(self, temperature: float) -> dict[str, Any]: - return {"temperature": temperature, "num_ctx": self.config.num_ctx} + return { + "temperature": temperature, + "num_ctx": self.config.num_ctx, + "seed": self.config.seed, + } def _require_structured_object(self, raw_response: str, schema: dict[str, Any], operation: str) -> dict[str, Any]: content = _parse_jsonish(raw_response) @@ -303,6 +319,7 @@ def health(self) -> dict[str, Any]: "active_embed_model": self.config.embed_model, "timeout_seconds": self.config.timeout_seconds, "num_ctx": self.config.num_ctx, + "seed": self.config.seed, "reachable": False, "status": "degraded", "error": str(e), @@ -327,6 +344,7 @@ def health(self) -> dict[str, Any]: "active_embed_model": active_embed_model, "timeout_seconds": self.config.timeout_seconds, "num_ctx": self.config.num_ctx, + "seed": self.config.seed, } return { **base, @@ -379,7 +397,7 @@ def summarize(self, text: str, title: str | None = None, url: str | None = None, "system": "You write precise summaries for selected text in a browser side panel.", "stream": False, "format": schema, - "options": self._generation_options(0.2), + "options": self._generation_options(0.0), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() @@ -434,7 +452,7 @@ def agent(self, prompt: str, context: dict[str, Any] | None = None) -> dict[str, "system": "You are a practical browser copilot that rewrites and structures selected text locally.", "stream": False, "format": schema, - "options": self._generation_options(0.2), + "options": self._generation_options(0.0), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() @@ -533,7 +551,7 @@ def extract( "system": "You generate clean structured extraction results for highlighted browser text.", "stream": False, "format": preset.schema, - "options": self._generation_options(0.1), + "options": self._generation_options(0.0), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() diff --git a/tests/server/test_installation_manager.py b/tests/server/test_installation_manager.py index 0fe2354..9db7dae 100644 --- a/tests/server/test_installation_manager.py +++ b/tests/server/test_installation_manager.py @@ -1,10 +1,11 @@ """module_name: installation_manager_tests; spec_ref: "testing_strategy.integration_tests".""" from __future__ import annotations +import json import sys import unittest from pathlib import Path -from unittest.mock import patch +from unittest.mock import MagicMock, patch ROOT = Path(__file__).resolve().parents[2] SERVER_DIR = ROOT / "server" @@ -34,6 +35,36 @@ def test_repeated_start_is_single_flight(self) -> None: state = manager.start(True) self.assertEqual(state["state"], "installing") + def test_warmup_uses_deterministic_profile_options(self) -> None: + response = MagicMock() + response.__enter__.return_value = response + response.__exit__.return_value = False + response.read.return_value = b'{"done":true}' + with patch("installation_manager.urlopen", return_value=response) as mocked: + InstallationManager()._warm_model("gemma4:e4b-it-qat", 32768) + request = mocked.call_args.args[0] + payload = json.loads(request.data) + self.assertEqual(payload["model"], "gemma4:e4b-it-qat") + self.assertEqual(payload["options"], { + "temperature": 0, + "seed": 42, + "num_ctx": 32768, + "num_predict": 8, + }) + + def test_warmup_uses_runtime_seed_override(self) -> None: + response = MagicMock() + response.__enter__.return_value = response + response.__exit__.return_value = False + response.read.return_value = b'{"done":true}' + with ( + patch.dict("installation_manager.os.environ", {"CHROMEAI_OLLAMA_SEED": "7"}), + patch("installation_manager.urlopen", return_value=response) as mocked, + ): + InstallationManager()._warm_model("gemma4:e4b-it-qat", 32768) + payload = json.loads(mocked.call_args.args[0].data) + self.assertEqual(payload["options"]["seed"], 7) + if __name__ == "__main__": unittest.main() diff --git a/tests/server/test_ollama_client.py b/tests/server/test_ollama_client.py index 49e4d84..9fe2fb2 100644 --- a/tests/server/test_ollama_client.py +++ b/tests/server/test_ollama_client.py @@ -22,6 +22,7 @@ def __init__(self, response: dict): embed_model="nomic-embed-text-v2-moe:latest", timeout_seconds=1, num_ctx=16384, + seed=42, )) self.response = response self.last_payload = None @@ -49,6 +50,8 @@ def test_summarize_sends_explicit_context_window(self) -> None: }) client.summarize("content") self.assertEqual(client.last_payload["options"]["num_ctx"], 16384) + self.assertEqual(client.last_payload["options"]["seed"], 42) + self.assertEqual(client.last_payload["options"]["temperature"], 0.0) def test_summarize_rejects_non_schema_output(self) -> None: client = StubOllamaClient({"response": "not json"}) From 834f158818019909296fa9ee0b2d43e3d77ef145 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 18:18:08 +0200 Subject: [PATCH 40/66] fix(runtime): support macOS system Python Add a Python 3.9 CI lane and defer runtime annotation evaluation so the local bridge runs on supported Apple system Python installations. Verified: pnpm lint && pnpm typecheck && pnpm build && pnpm test Verified: pnpm test:privacy && pnpm test:e2e Verified: node scripts/stress-runner.mjs && pnpm bench:frontier Verified: pnpm audit --audit-level=moderate Verified: physical MacBook Air M1 Python 3.9 bridge, deterministic extraction, fail-closed recovery --- .github/workflows/ci.yml | 22 ++++++++++++++++++++++ server/nano_server.py | 2 ++ server/ollama_client.py | 2 ++ 3 files changed, 26 insertions(+) diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 5514067..fd179e1 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -11,6 +11,28 @@ concurrency: cancel-in-progress: true jobs: + python-runtime-compat: + runs-on: ubuntu-latest + permissions: + contents: read + + steps: + - uses: actions/checkout@v7 + + - name: Use Python 3.9 + uses: actions/setup-python@v7 + with: + python-version: "3.9" + + - name: Install local billing runtime dependencies + run: python -m pip install --requirement requirements-billing.txt + + - name: Smoke macOS system-Python bridge import + run: python server/nano_server.py --help + + - name: Test server on Python 3.9 + run: python -m unittest discover -s tests/server -p 'test_*.py' + validate: runs-on: ubuntu-latest permissions: diff --git a/server/nano_server.py b/server/nano_server.py index 6969784..0fa7a13 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -5,6 +5,8 @@ # @spec_ref model_switch_hysteresis # @spec_ref failure_isolation # @spec_ref latency_budget +from __future__ import annotations + import argparse import hashlib import json diff --git a/server/ollama_client.py b/server/ollama_client.py index 98cc076..0abf54b 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -1,6 +1,8 @@ # module_name: ollama_runtime # spec_ref: "provisioning_layer" # @spec_ref prompt_determinism +from __future__ import annotations + import json import os import re From edc62f89a4056471a3ca2b540a7920318cccc7a2 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 18:24:18 +0200 Subject: [PATCH 41/66] fix(runtime): remove incompatible HTTP dependency Use the Python standard library for the loopback-only Monero RPC membrane so supported macOS Python 3.9 installations do not require an incompatible requests release. Verified: Python 3.9 full 63-test server suite Verified: pnpm lint && pnpm typecheck && pnpm build && pnpm test Verified: pnpm test:privacy && pnpm test:e2e Verified: node scripts/stress-runner.mjs && pnpm bench:frontier Verified: pnpm audit --audit-level=moderate --- requirements-billing.txt | 1 - server/billing_security.py | 24 ++++++++++++++++++++++++ server/monero_payment_server.py | 25 ++++++++----------------- tests/server/test_billing_security.py | 24 ++++++++++++++++++++++++ 4 files changed, 56 insertions(+), 18 deletions(-) diff --git a/requirements-billing.txt b/requirements-billing.txt index 4feeaac..af266d2 100644 --- a/requirements-billing.txt +++ b/requirements-billing.txt @@ -1,2 +1 @@ Flask==3.1.3 -requests==2.34.2 diff --git a/server/billing_security.py b/server/billing_security.py index b19ad24..303423a 100644 --- a/server/billing_security.py +++ b/server/billing_security.py @@ -8,6 +8,7 @@ import secrets from pathlib import Path from urllib.parse import urlparse +from urllib.request import Request, urlopen LOCAL_HOSTS = {"127.0.0.1", "localhost", "::1"} @@ -20,6 +21,29 @@ def validated_wallet_rpc_url(value: str) -> str: return value +def call_wallet_rpc(url: str, method: str, params: dict | None = None) -> dict: + endpoint = validated_wallet_rpc_url(url) + body = json.dumps( + { + "jsonrpc": "2.0", + "id": "0", + "method": method, + "params": params or {}, + } + ).encode("utf-8") + rpc_request = Request( + endpoint, + data=body, + headers={"Content-Type": "application/json"}, + method="POST", + ) + with urlopen(rpc_request, timeout=10) as response: + payload = json.load(response) + if "error" in payload: + raise RuntimeError(f"Monero RPC error: {payload['error']}") + return payload["result"] + + def new_order_id() -> str: return "SP-" + secrets.token_hex(16) diff --git a/server/monero_payment_server.py b/server/monero_payment_server.py index dab1316..b7b6dd2 100644 --- a/server/monero_payment_server.py +++ b/server/monero_payment_server.py @@ -21,9 +21,14 @@ from pathlib import Path from threading import Lock, Thread -import requests from flask import Flask, jsonify, request -from billing_security import admin_secret_matches, new_order_id, save_private_json, validated_wallet_rpc_url +from billing_security import ( + admin_secret_matches, + call_wallet_rpc, + new_order_id, + save_private_json, + validated_wallet_rpc_url, +) from entitlement_signer import EntitlementSigner, SigningError app = Flask(__name__) @@ -57,21 +62,7 @@ def save_db(db): # ---- RPC ---- def rpc(method, params=None): - r = requests.post( - RPC_URL, - json={ - "jsonrpc": "2.0", - "id": "0", - "method": method, - "params": params or {}, - }, - timeout=10, - ) - r.raise_for_status() - payload = r.json() - if "error" in payload: - raise RuntimeError(f"Monero RPC error: {payload['error']}") - return payload["result"] + return call_wallet_rpc(RPC_URL, method, params) def xmr_to_atomic(xmr_amount: float) -> int: diff --git a/tests/server/test_billing_security.py b/tests/server/test_billing_security.py index 08d85c5..f5eee86 100644 --- a/tests/server/test_billing_security.py +++ b/tests/server/test_billing_security.py @@ -2,12 +2,15 @@ from __future__ import annotations +import io +import json import os import stat import sys import tempfile import unittest from pathlib import Path +from unittest.mock import patch ROOT = Path(__file__).resolve().parents[2] SERVER_DIR = ROOT / "server" @@ -16,6 +19,7 @@ from billing_security import ( # noqa: E402 admin_secret_matches, + call_wallet_rpc, new_order_id, save_private_json, validated_wallet_rpc_url, @@ -23,6 +27,26 @@ class BillingSecurityTests(unittest.TestCase): + def test_wallet_rpc_uses_loopback_standard_library_request(self) -> None: + response = io.BytesIO( + json.dumps({"jsonrpc": "2.0", "id": "0", "result": {"height": 42}}).encode() + ) + + with patch("billing_security.urlopen", return_value=response) as request_mock: + self.assertEqual( + call_wallet_rpc("http://127.0.0.1:18083/json_rpc", "get_height"), + {"height": 42}, + ) + + sent_request = request_mock.call_args.args[0] + self.assertEqual(sent_request.method, "POST") + self.assertEqual(sent_request.full_url, "http://127.0.0.1:18083/json_rpc") + self.assertEqual( + json.loads(sent_request.data), + {"jsonrpc": "2.0", "id": "0", "method": "get_height", "params": {}}, + ) + self.assertEqual(request_mock.call_args.kwargs["timeout"], 10) + def test_wallet_rpc_is_loopback_only(self) -> None: self.assertEqual( validated_wallet_rpc_url("http://127.0.0.1:18083/json_rpc"), From 6533c3ffc6382271d6d337d0e46991e4e4722f01 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 19:30:59 +0200 Subject: [PATCH 42/66] feat(runtime): route tasks to qualified local models Bind structured tasks to Gemma E2B and agent tasks to Gemma E4B, enforce a deterministic local input budget, and provision the exact model bundle without fallback substitution. Regenerate fail-closed runtime policy evidence from the qualified model registry.\n\nVerified: Hetzner lint, typecheck, build, 25 panel tests, 68 server tests, privacy, 7/7 E2E, policy validation, 10/10 stress, Store package; Python 3.9 server/billing; physical M1 route/failure/recovery acceptance. --- README.md | 2 +- docs/model-runtime-decisions.md | 4 + launchd/com.chromeai.nano.plist | 3 + panel/runtime-profiles.js | 13 +- panel/runtime-profiles.ts | 16 +- reports/aggregated_metrics.json | 28 +- reports/architecture_decision_report.json | 2 +- reports/bottleneck_analysis.json | 11 +- reports/determinism_report.json | 6 +- reports/engine_decision_report.json | 8 +- reports/execution_matrix.json | 2688 ++---- reports/frontier_decisions.json | 44 +- reports/frontier_report.json | 2 +- reports/frontier_summary.html | 6 +- reports/model_comparison.json | 34 +- reports/phase_report.json | 20 +- reports/raw_metrics.jsonl | 6720 +++++--------- reports/repo_plan.json | 2 +- reports/runtime_policy_validation.json | 4 +- reports/runtime_selection_summary.json | 7 +- reports/scenario_results.json | 7804 ++++------------- reports/simulation_report.json | 2 +- reports/spec_coverage_report.json | 2 +- reports/verification_report.json | 2 +- runtime/model_policy.json | 97 +- runtime/model_registry.runtime.json | 55 +- runtime/promotion_audit.json | 61 +- scripts/bootstrap-macos-local.sh | 84 +- scripts/install-macos-local.sh | 19 +- scripts/runtime-policy-lib.mjs | 39 +- scripts/spec-compiler.mjs | 4 +- selectpilot_benchmark_v1.json | 8 +- server/installation_manager.py | 16 +- server/model/registry_source.json | 31 +- server/nano_server.py | 80 +- server/ollama_client.py | 90 +- server/runtime_profiles.py | 49 +- tests/panel/bootstrap-contract.test.mjs | 10 + .../panel/runtime-policy-determinism.test.mjs | 19 + tests/server/test_ollama_client.py | 41 +- tests/server/test_runtime_profiles.py | 22 +- tests/server/test_validation_pipeline.py | 14 + 42 files changed, 5035 insertions(+), 13134 deletions(-) create mode 100644 tests/panel/runtime-policy-determinism.test.mjs diff --git a/README.md b/README.md index 179522d..715aa33 100644 --- a/README.md +++ b/README.md @@ -47,7 +47,7 @@ SelectPilot is built to keep the core workflow local-first and inspectable. pnpm setup:local ``` -This installs dependencies, builds the extension, installs/starts Ollama when needed, selects a hardware-safe local profile, pulls and prewarms the exact generation model, installs the LaunchAgent, and refuses to finish unless the bridge and both configured models are healthy. Structured generation uses a fixed zero-temperature sampling contract and seed so repeated requests follow the same decision path. +This installs dependencies, builds the extension, installs/starts Ollama when needed, selects a hardware-safe local profile, prepares the exact local model bundle, installs the LaunchAgent, and refuses to finish unless the bridge and configured models are healthy. Extraction and summaries use the smallest qualified local model; broader agent tasks use the profile's general model. Structured generation uses a fixed zero-temperature sampling contract and seed so repeated requests follow the same decision path. 2. Open `chrome://extensions`, enable **Developer mode**, choose **Load unpacked**, and select this repository folder. 3. Highlight text on a page → open SelectPilot → click **Extract JSON**. diff --git a/docs/model-runtime-decisions.md b/docs/model-runtime-decisions.md index 59abd14..cf92dfd 100644 --- a/docs/model-runtime-decisions.md +++ b/docs/model-runtime-decisions.md @@ -8,6 +8,10 @@ - Embeddings: `nomic-embed-text-v2-moe:latest`. - Structured generation: `temperature=0` and `seed=42` for every production profile. +Structured extraction and summaries use `gemma4:e2b-it-qat`, the smallest model qualified on the physical M1 acceptance machine. The Balanced profile keeps `gemma4:e4b-it-qat` for the more open-ended agent route. Every route declares its exact model and context window in bridge health; a missing routed model stops that operation rather than substituting another model. + +Selections above 16,000 characters are rejected before inference with a clear, local error. This keeps waiting time predictable and prevents hidden truncation. + Explicit environment overrides remain authoritative and must fail visibly when invalid or unavailable. SelectPilot never substitutes a different installed model, downloads a model without consent, or routes generation to a cloud model. Installation prewarms the selected generation model with the same context window and deterministic sampling options used at runtime. A failed prewarm is an explicit installation failure; SelectPilot does not report the runtime as ready and does not silently choose another model. `CHROMEAI_OLLAMA_SEED` may override the seed for controlled verification, but invalid or negative values fail closed. diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index 45c1b71..4b4de4e 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -24,8 +24,11 @@ CHROMEAI_ALLOWED_RANGE8083 CHROMEAI_OLLAMA_BASE_URL__OLLAMA_BASE_URL__ CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ + CHROMEAI_OLLAMA_FAST_MODEL__OLLAMA_FAST_MODEL__ CHROMEAI_OLLAMA_EMBED_MODEL__OLLAMA_EMBED_MODEL__ CHROMEAI_OLLAMA_NUM_CTX__OLLAMA_NUM_CTX__ + CHROMEAI_OLLAMA_FAST_NUM_CTX__OLLAMA_FAST_NUM_CTX__ + CHROMEAI_MAX_INPUT_CHARS__MAX_INPUT_CHARS__ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_AUTO_PULL_MODELS0 diff --git a/panel/runtime-profiles.js b/panel/runtime-profiles.js index 52f48ec..be37ae8 100644 --- a/panel/runtime-profiles.js +++ b/panel/runtime-profiles.js @@ -4,9 +4,12 @@ export const RUNTIME_PROFILES = [ label: 'Fast', description: 'Smallest viable local profile for structured extraction and low-latency summaries.', generation_model: 'gemma4:e2b-it-qat', + fast_generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 16384, - target_latency: '1-4s', + fast_num_ctx: 16384, + max_input_chars: 16000, + target_latency: '4-20s', intended_for: 'Selected-text extraction, action briefs, and quick summaries.', command: './scripts/bootstrap-macos-local.sh --profile fast', is_default_auto: true, @@ -16,9 +19,12 @@ export const RUNTIME_PROFILES = [ label: 'Balanced', description: 'Higher quality local profile for rewrite and general-purpose browser transforms.', generation_model: 'gemma4:e4b-it-qat', + fast_generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 32768, - target_latency: '2-6s', + fast_num_ctx: 16384, + max_input_chars: 16000, + target_latency: '4-30s', intended_for: 'Daily use when you want better quality without drifting into heavy models.', command: './scripts/bootstrap-macos-local.sh --profile balanced', }, @@ -27,8 +33,11 @@ export const RUNTIME_PROFILES = [ label: 'Advanced', description: 'Manual opt-in profile for stronger reasoning on larger machines.', generation_model: 'qwen2.5:7b', + fast_generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 32768, + fast_num_ctx: 16384, + max_input_chars: 16000, target_latency: '4-10s', intended_for: 'Heavier rewrite and ask flows when latency budget is less important.', command: './scripts/bootstrap-macos-local.sh --profile advanced', diff --git a/panel/runtime-profiles.ts b/panel/runtime-profiles.ts index 23aa38a..f072831 100644 --- a/panel/runtime-profiles.ts +++ b/panel/runtime-profiles.ts @@ -7,8 +7,11 @@ export type RuntimeProfile = { label: string; description: string; generation_model: string; + fast_generation_model: string; embedding_model: string; num_ctx: number; + fast_num_ctx: number; + max_input_chars: number; target_latency: string; intended_for: string; command: string; @@ -21,9 +24,12 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ label: 'Fast', description: 'Smallest viable local profile for structured extraction and low-latency summaries.', generation_model: 'gemma4:e2b-it-qat', + fast_generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 16384, - target_latency: '1-4s', + fast_num_ctx: 16384, + max_input_chars: 16000, + target_latency: '4-20s', intended_for: 'Selected-text extraction, action briefs, and quick summaries.', command: './scripts/bootstrap-macos-local.sh --profile fast', is_default_auto: true, @@ -33,9 +39,12 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ label: 'Balanced', description: 'Higher quality local profile for rewrite and general-purpose browser transforms.', generation_model: 'gemma4:e4b-it-qat', + fast_generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 32768, - target_latency: '2-6s', + fast_num_ctx: 16384, + max_input_chars: 16000, + target_latency: '4-30s', intended_for: 'Daily use when you want better quality without drifting into heavy models.', command: './scripts/bootstrap-macos-local.sh --profile balanced', }, @@ -44,8 +53,11 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ label: 'Advanced', description: 'Manual opt-in profile for stronger reasoning on larger machines.', generation_model: 'qwen2.5:7b', + fast_generation_model: 'gemma4:e2b-it-qat', embedding_model: 'nomic-embed-text-v2-moe:latest', num_ctx: 32768, + fast_num_ctx: 16384, + max_input_chars: 16000, target_latency: '4-10s', intended_for: 'Heavier rewrite and ask flows when latency budget is less important.', command: './scripts/bootstrap-macos-local.sh --profile advanced', diff --git a/reports/aggregated_metrics.json b/reports/aggregated_metrics.json index ac1361b..e8cd467 100644 --- a/reports/aggregated_metrics.json +++ b/reports/aggregated_metrics.json @@ -3,30 +3,30 @@ "runtime_verified": false, "promotion_eligible": false, "totals": { - "execution_matrix_rows": 448, - "runs": 4480, + "execution_matrix_rows": 224, + "runs": 2240, "runs_per_matrix_row": 10 }, "latency": { - "p50_latency_ms": 951, - "p95_latency_ms": 3014, - "p99_latency_ms": 4000 + "p50_latency_ms": 679, + "p95_latency_ms": 1505, + "p99_latency_ms": 1872 }, "correctness": { - "task_success_rate": 0.9314, - "schema_validity_rate": 0.9688, + "task_success_rate": 0.9396, + "schema_validity_rate": 0.9709, "retry_rate": 0.0593, "fallback_rate": 0.0157 }, "stability": { - "failure_rate": 0.0568, - "validation_failure_rate": 0.0387, - "quarantine_trigger_rate": 0.0245, - "oscillation_rate": 0.0399 + "failure_rate": 0.0493, + "validation_failure_rate": 0.0334, + "quarantine_trigger_rate": 0.0236, + "oscillation_rate": 0.0401 }, "meta_metrics": { - "determinism_score": 0.9803, - "variance_across_runs": 0.0199, - "confidence_interval": 0.94 + "determinism_score": 0.9804, + "variance_across_runs": 0.02, + "confidence_interval": 0.9402 } } diff --git a/reports/architecture_decision_report.json b/reports/architecture_decision_report.json index cc8375e..39557be 100644 --- a/reports/architecture_decision_report.json +++ b/reports/architecture_decision_report.json @@ -2,7 +2,7 @@ "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "generated_at": "2026-08-18T19:23:33.758Z", + "generated_at": "2026-08-25T17:30:23.119Z", "engine_recommendation": "Do not start with full C++ engine. Start with high-level orchestration and instrument everything. Add native hot paths only if measurements justify it.", "bottleneck_summary": [ "strict_json_validation", diff --git a/reports/bottleneck_analysis.json b/reports/bottleneck_analysis.json index c6c8ad8..be8ab49 100644 --- a/reports/bottleneck_analysis.json +++ b/reports/bottleneck_analysis.json @@ -3,13 +3,12 @@ "runtime_verified": false, "promotion_eligible": false, "detected": [ - "model_latency_dominance", "memory_pressure_events" ], "summary": "bottlenecks_detected", - "p95_latency_ms": 3014, - "orchestration_overhead_ratio": 0.056, - "validation_overhead_ratio": 0.0435, - "event_dispatch_cost": 8.1257, - "memory_management_cost": 8.8977 + "p95_latency_ms": 1505, + "orchestration_overhead_ratio": 0.0909, + "validation_overhead_ratio": 0.0592, + "event_dispatch_cost": 6.6657, + "memory_management_cost": 8.7236 } diff --git a/reports/determinism_report.json b/reports/determinism_report.json index 0ff0150..122703c 100644 --- a/reports/determinism_report.json +++ b/reports/determinism_report.json @@ -3,10 +3,10 @@ "runtime_verified": false, "promotion_eligible": false, "target": 0.95, - "group_count": 448, - "bounded_output_variance": 0.9803, + "group_count": 224, + "bounded_output_variance": 0.9804, "same_model_selection": true, "no_random_fallbacks": true, - "score": 0.9803, + "score": 0.9804, "status": "pass" } diff --git a/reports/engine_decision_report.json b/reports/engine_decision_report.json index 46ace5c..c4071e5 100644 --- a/reports/engine_decision_report.json +++ b/reports/engine_decision_report.json @@ -11,9 +11,9 @@ "memory_management_cost_max": 15 }, "input": { - "orchestration_overhead_ratio": 0.056, - "validation_overhead_ratio": 0.0435, - "event_dispatch_cost": 8.1257, - "memory_management_cost": 8.8977 + "orchestration_overhead_ratio": 0.0909, + "validation_overhead_ratio": 0.0592, + "event_dispatch_cost": 6.6657, + "memory_management_cost": 8.7236 } } diff --git a/reports/execution_matrix.json b/reports/execution_matrix.json index f14f143..585b199 100644 --- a/reports/execution_matrix.json +++ b/reports/execution_matrix.json @@ -2,3136 +2,1568 @@ { "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:7b", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:7b", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:14b", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:14b", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:7b", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:7b", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:14b", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:14b", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "classification_binary", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "classification_binary", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "classification_binary", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "classification_binary", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", + "scenario": "classification_binary", "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:7b", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:7b", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:14b", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:14b", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:0.5b", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:0.5b", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:1.5b", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:1.5b", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "classification_multiclass", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "classification_multiclass", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "classification_multiclass", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "classification_multiclass", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "extract_entities", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "extract_entities", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "extract_entities", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "extract_entities", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", + "scenario": "extract_entities", "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold" - }, - { - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm" - }, - { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "compare_structured", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "ambiguous_intent", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "ambiguous_intent", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "analyze_reasoning", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "analyze_reasoning", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "analyze_reasoning", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "analyze_reasoning", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "analyze_reasoning", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "contradictory_intent", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "contradictory_intent", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "long_context", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "long_context", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:7b", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:14b", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:7b", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:14b", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" }, { "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold" }, { "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm" } diff --git a/reports/frontier_decisions.json b/reports/frontier_decisions.json index b75ac04..dbd72ec 100644 --- a/reports/frontier_decisions.json +++ b/reports/frontier_decisions.json @@ -3,52 +3,26 @@ "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "baseline_model": "qwen2:1.5b", - "candidate_model": "qwen2:0.5b", - "decision": "promote", - "reason": "higher_correctness_same_or_lower_cost", - "latency_regression_ratio": 0.7612, - "schema_validity_rate": 0.9717, - "failure_rate": 0.0493, - "weighted_score": 0.8854 - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "baseline_model": "qwen2:1.5b", - "candidate_model": "qwen2:1.5b", + "baseline_model": "gemma4:e2b-it-qat", + "candidate_model": "gemma4:e2b-it-qat", "decision": "reject", "reason": "higher_failure_rate", "latency_regression_ratio": 1, - "schema_validity_rate": 0.9709, + "schema_validity_rate": 0.9717, "failure_rate": 0.0493, - "weighted_score": 0.857 + "weighted_score": 0.8822 }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "baseline_model": "qwen2:1.5b", - "candidate_model": "qwen2:7b", + "baseline_model": "gemma4:e2b-it-qat", + "candidate_model": "gemma4:e4b-it-qat", "decision": "reject", "reason": "increases_latency_without_gain", - "latency_regression_ratio": 1.9176, - "schema_validity_rate": 0.9678, + "latency_regression_ratio": 1.6184, + "schema_validity_rate": 0.9701, "failure_rate": 0.0493, - "weighted_score": 0.7605 - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "baseline_model": "qwen2:1.5b", - "candidate_model": "qwen2:14b", - "decision": "reject", - "reason": "increases_latency_without_gain", - "latency_regression_ratio": 2.8, - "schema_validity_rate": 0.9648, - "failure_rate": 0.0793, - "weighted_score": 0.6719 + "weighted_score": 0.8334 } ] diff --git a/reports/frontier_report.json b/reports/frontier_report.json index 2d46a04..9bd76ce 100644 --- a/reports/frontier_report.json +++ b/reports/frontier_report.json @@ -2,7 +2,7 @@ "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "generated_at": "2026-08-18T19:23:33.758Z", + "generated_at": "2026-08-25T17:30:23.119Z", "current_frontier": [], "candidate_comparisons": [], "promotion_decisions": [], diff --git a/reports/frontier_summary.html b/reports/frontier_summary.html index 05404c1..53453eb 100644 --- a/reports/frontier_summary.html +++ b/reports/frontier_summary.html @@ -19,10 +19,8 @@

    Frontier Decisions

    - qwen2:0.5bpromotehigher_correctness_same_or_lower_cost0.88540.7612 -qwen2:1.5brejecthigher_failure_rate0.8571 -qwen2:7brejectincreases_latency_without_gain0.76051.9176 -qwen2:14brejectincreases_latency_without_gain0.67192.8 + gemma4:e2b-it-qatrejecthigher_failure_rate0.88221 +gemma4:e4b-it-qatrejectincreases_latency_without_gain0.83341.6184 diff --git a/reports/model_comparison.json b/reports/model_comparison.json index 2de550e..0ece01c 100644 --- a/reports/model_comparison.json +++ b/reports/model_comparison.json @@ -3,40 +3,20 @@ "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "model": "qwen2:0.5b", - "weighted_score": 0.8854, + "model": "gemma4:e2b-it-qat", + "weighted_score": 0.8822, "schema_validity_rate": 0.9717, "failure_rate": 0.0493, - "p95_latency_ms": 628 + "p95_latency_ms": 629 }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "model": "qwen2:1.5b", - "weighted_score": 0.857, - "schema_validity_rate": 0.9709, + "model": "gemma4:e4b-it-qat", + "weighted_score": 0.8334, + "schema_validity_rate": 0.9701, "failure_rate": 0.0493, - "p95_latency_ms": 825 - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "model": "qwen2:7b", - "weighted_score": 0.7605, - "schema_validity_rate": 0.9678, - "failure_rate": 0.0493, - "p95_latency_ms": 1582 - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "model": "qwen2:14b", - "weighted_score": 0.6719, - "schema_validity_rate": 0.9648, - "failure_rate": 0.0793, - "p95_latency_ms": 2310 + "p95_latency_ms": 1018 } ] diff --git a/reports/phase_report.json b/reports/phase_report.json index 8731dc5..20fbad9 100644 --- a/reports/phase_report.json +++ b/reports/phase_report.json @@ -8,7 +8,7 @@ "missing_node_report" ], "completion_score": 2, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_1_repo_plan", @@ -19,7 +19,7 @@ "module_manifest.json" ], "completion_score": 34, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_2_type_contracts", @@ -32,7 +32,7 @@ "shared/types/events.ts" ], "completion_score": 4, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_3_module_scaffolds", @@ -42,7 +42,7 @@ "scaffold_report.json" ], "completion_score": 1, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_4_core_implementation", @@ -50,7 +50,7 @@ "objective": "Implement backend deterministic core before UI polish.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_5_ui_implementation", @@ -58,7 +58,7 @@ "objective": "Implement frontend surfaces after runtime contracts are stable.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_6_verification", @@ -66,7 +66,7 @@ "objective": "Run static, behavioral, privacy, and state-machine validation.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_7_simulation", @@ -74,7 +74,7 @@ "objective": "Run realistic local simulations against multiple models, hardware profiles, and task scenarios.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_8_benchmark_analysis", @@ -82,7 +82,7 @@ "objective": "Compute metrics, frontier comparisons, regressions, and promotion eligibility.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" }, { "phase_id": "phase_9_decision_reports", @@ -90,6 +90,6 @@ "objective": "Emit actionable reports on model policy, performance, and architecture choices.", "outputs_declared": [], "completion_score": 0, - "generated_at": "2026-08-18T19:23:33.758Z" + "generated_at": "2026-08-25T17:30:23.118Z" } ] diff --git a/reports/raw_metrics.jsonl b/reports/raw_metrics.jsonl index 4b9e534..8248821 100644 --- a/reports/raw_metrics.jsonl +++ b/reports/raw_metrics.jsonl @@ -1,4480 +1,2240 @@ -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":822,"p95_latency_ms":1005,"p99_latency_ms":1078,"time_to_first_token_ms":205,"total_runtime_ms":914,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.013,"confidence_interval":0.9368,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7788,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":851,"p95_latency_ms":1040,"p99_latency_ms":1115,"time_to_first_token_ms":206,"total_runtime_ms":945,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0299,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0703,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1055,"time_to_first_token_ms":200,"total_runtime_ms":894,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0133,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9361,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":842,"p95_latency_ms":1029,"p99_latency_ms":1104,"time_to_first_token_ms":211,"total_runtime_ms":936,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0276,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3191,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":810,"p95_latency_ms":990,"p99_latency_ms":1063,"time_to_first_token_ms":204,"total_runtime_ms":900,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0151,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1166,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":788,"p95_latency_ms":963,"p99_latency_ms":1033,"time_to_first_token_ms":223,"total_runtime_ms":876,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0223,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9602,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":855,"p95_latency_ms":1045,"p99_latency_ms":1121,"time_to_first_token_ms":198,"total_runtime_ms":950,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0173,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4022,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":762,"p95_latency_ms":932,"p99_latency_ms":1000,"time_to_first_token_ms":220,"total_runtime_ms":847,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0245,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3153,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":780,"p95_latency_ms":953,"p99_latency_ms":1022,"time_to_first_token_ms":209,"total_runtime_ms":866,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0217,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0482,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":814,"p95_latency_ms":995,"p99_latency_ms":1068,"time_to_first_token_ms":192,"total_runtime_ms":905,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0183,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9116,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":385,"p95_latency_ms":471,"p99_latency_ms":505,"time_to_first_token_ms":151,"total_runtime_ms":428,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0239,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5281,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":383,"p95_latency_ms":469,"p99_latency_ms":503,"time_to_first_token_ms":174,"total_runtime_ms":426,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.019,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9153,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":176,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0216,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4235,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":458,"p95_latency_ms":560,"p99_latency_ms":601,"time_to_first_token_ms":162,"total_runtime_ms":509,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0142,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7826,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":465,"p95_latency_ms":568,"p99_latency_ms":610,"time_to_first_token_ms":149,"total_runtime_ms":517,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0275,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1976,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":395,"p95_latency_ms":482,"p99_latency_ms":518,"time_to_first_token_ms":144,"total_runtime_ms":439,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.904,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":437,"p95_latency_ms":534,"p99_latency_ms":573,"time_to_first_token_ms":154,"total_runtime_ms":486,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0205,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9948,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":457,"p95_latency_ms":558,"p99_latency_ms":599,"time_to_first_token_ms":168,"total_runtime_ms":507,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0145,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7958,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":426,"p95_latency_ms":521,"p99_latency_ms":559,"time_to_first_token_ms":165,"total_runtime_ms":473,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0187,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6501,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":477,"p95_latency_ms":583,"p99_latency_ms":625,"time_to_first_token_ms":176,"total_runtime_ms":530,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0223,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1207,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1062,"p95_latency_ms":1298,"p99_latency_ms":1393,"time_to_first_token_ms":235,"total_runtime_ms":1180,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0188,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6607,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1021,"p95_latency_ms":1247,"p99_latency_ms":1338,"time_to_first_token_ms":237,"total_runtime_ms":1134,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0243,"confidence_interval":0.9447,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.349,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1028,"p95_latency_ms":1257,"p99_latency_ms":1348,"time_to_first_token_ms":263,"total_runtime_ms":1142,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0149,"confidence_interval":0.9385,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3901,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1105,"p95_latency_ms":1350,"p99_latency_ms":1448,"time_to_first_token_ms":258,"total_runtime_ms":1228,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0125,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5652,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1100,"p95_latency_ms":1345,"p99_latency_ms":1442,"time_to_first_token_ms":261,"total_runtime_ms":1222,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0124,"confidence_interval":0.9471,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6397,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1076,"p95_latency_ms":1315,"p99_latency_ms":1411,"time_to_first_token_ms":266,"total_runtime_ms":1196,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0156,"confidence_interval":0.9416,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3159,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1100,"p95_latency_ms":1345,"p99_latency_ms":1442,"time_to_first_token_ms":241,"total_runtime_ms":1222,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0122,"confidence_interval":0.932,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9222,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1031,"p95_latency_ms":1260,"p99_latency_ms":1352,"time_to_first_token_ms":264,"total_runtime_ms":1145,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0265,"confidence_interval":0.9315,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1398,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1110,"p95_latency_ms":1357,"p99_latency_ms":1455,"time_to_first_token_ms":248,"total_runtime_ms":1233,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0232,"confidence_interval":0.9423,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9131,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1027,"p95_latency_ms":1256,"p99_latency_ms":1347,"time_to_first_token_ms":261,"total_runtime_ms":1141,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0165,"confidence_interval":0.95,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5109,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":552,"p95_latency_ms":675,"p99_latency_ms":724,"time_to_first_token_ms":202,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0108,"confidence_interval":0.9482,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6463,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":508,"p95_latency_ms":621,"p99_latency_ms":666,"time_to_first_token_ms":219,"total_runtime_ms":564,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0185,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7552,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":553,"p95_latency_ms":676,"p99_latency_ms":725,"time_to_first_token_ms":212,"total_runtime_ms":615,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0136,"confidence_interval":0.937,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9847,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":557,"p95_latency_ms":680,"p99_latency_ms":730,"time_to_first_token_ms":195,"total_runtime_ms":619,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0193,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8717,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":561,"p95_latency_ms":686,"p99_latency_ms":736,"time_to_first_token_ms":220,"total_runtime_ms":624,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0122,"confidence_interval":0.936,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5667,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":598,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":214,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.028,"confidence_interval":0.944,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1966,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":566,"p95_latency_ms":692,"p99_latency_ms":742,"time_to_first_token_ms":193,"total_runtime_ms":629,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0171,"confidence_interval":0.9378,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9184,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":214,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0101,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4825,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":540,"p95_latency_ms":660,"p99_latency_ms":708,"time_to_first_token_ms":205,"total_runtime_ms":600,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0215,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7511,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":573,"p95_latency_ms":701,"p99_latency_ms":752,"time_to_first_token_ms":207,"total_runtime_ms":637,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0196,"confidence_interval":0.9424,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1631,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2067,"p95_latency_ms":2526,"p99_latency_ms":2710,"time_to_first_token_ms":427,"total_runtime_ms":2297,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0162,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5345,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2111,"p95_latency_ms":2580,"p99_latency_ms":2767,"time_to_first_token_ms":413,"total_runtime_ms":2345,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.9414,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9668,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2025,"p95_latency_ms":2475,"p99_latency_ms":2655,"time_to_first_token_ms":415,"total_runtime_ms":2250,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0116,"confidence_interval":0.9347,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0381,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2030,"p95_latency_ms":2481,"p99_latency_ms":2662,"time_to_first_token_ms":439,"total_runtime_ms":2256,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0288,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6363,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2046,"p95_latency_ms":2501,"p99_latency_ms":2683,"time_to_first_token_ms":431,"total_runtime_ms":2274,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0192,"confidence_interval":0.9474,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6516,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2084,"p95_latency_ms":2547,"p99_latency_ms":2733,"time_to_first_token_ms":421,"total_runtime_ms":2316,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0119,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1258,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2054,"p95_latency_ms":2511,"p99_latency_ms":2693,"time_to_first_token_ms":420,"total_runtime_ms":2282,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0272,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1025,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2029,"p95_latency_ms":2480,"p99_latency_ms":2660,"time_to_first_token_ms":438,"total_runtime_ms":2254,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0192,"confidence_interval":0.9412,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8981,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2071,"p95_latency_ms":2531,"p99_latency_ms":2715,"time_to_first_token_ms":427,"total_runtime_ms":2301,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0159,"confidence_interval":0.9443,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2517,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2093,"p95_latency_ms":2558,"p99_latency_ms":2744,"time_to_first_token_ms":439,"total_runtime_ms":2325,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0166,"confidence_interval":0.9347,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9641,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1051,"p95_latency_ms":1285,"p99_latency_ms":1378,"time_to_first_token_ms":393,"total_runtime_ms":1168,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0295,"confidence_interval":0.9408,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2837,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1094,"p95_latency_ms":1337,"p99_latency_ms":1434,"time_to_first_token_ms":380,"total_runtime_ms":1215,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0229,"confidence_interval":0.9388,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5146,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1067,"p95_latency_ms":1304,"p99_latency_ms":1399,"time_to_first_token_ms":394,"total_runtime_ms":1186,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0212,"confidence_interval":0.95,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8947,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1107,"p95_latency_ms":1353,"p99_latency_ms":1451,"time_to_first_token_ms":361,"total_runtime_ms":1230,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.02,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8276,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1019,"p95_latency_ms":1245,"p99_latency_ms":1336,"time_to_first_token_ms":372,"total_runtime_ms":1132,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0265,"confidence_interval":0.9316,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6101,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1042,"p95_latency_ms":1273,"p99_latency_ms":1366,"time_to_first_token_ms":394,"total_runtime_ms":1157,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0264,"confidence_interval":0.9326,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6329,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1106,"p95_latency_ms":1351,"p99_latency_ms":1450,"time_to_first_token_ms":392,"total_runtime_ms":1229,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0122,"confidence_interval":0.9305,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1835,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1112,"p95_latency_ms":1359,"p99_latency_ms":1458,"time_to_first_token_ms":365,"total_runtime_ms":1236,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0198,"confidence_interval":0.946,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0299,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1045,"p95_latency_ms":1277,"p99_latency_ms":1370,"time_to_first_token_ms":363,"total_runtime_ms":1161,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0128,"confidence_interval":0.939,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1251,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1091,"p95_latency_ms":1333,"p99_latency_ms":1430,"time_to_first_token_ms":364,"total_runtime_ms":1212,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0172,"confidence_interval":0.9467,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5182,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3061,"p95_latency_ms":3741,"p99_latency_ms":4013,"time_to_first_token_ms":594,"total_runtime_ms":3401,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0256,"confidence_interval":0.9358,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9522,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3010,"p95_latency_ms":3679,"p99_latency_ms":3946,"time_to_first_token_ms":585,"total_runtime_ms":3344,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.9384,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0769,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2996,"p95_latency_ms":3662,"p99_latency_ms":3928,"time_to_first_token_ms":604,"total_runtime_ms":3329,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0208,"confidence_interval":0.9409,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4732,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3093,"p95_latency_ms":3781,"p99_latency_ms":4056,"time_to_first_token_ms":587,"total_runtime_ms":3437,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0188,"confidence_interval":0.9467,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0379,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3076,"p95_latency_ms":3760,"p99_latency_ms":4033,"time_to_first_token_ms":588,"total_runtime_ms":3418,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0288,"confidence_interval":0.9461,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8177,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3010,"p95_latency_ms":3679,"p99_latency_ms":3947,"time_to_first_token_ms":597,"total_runtime_ms":3345,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0129,"confidence_interval":0.9407,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7102,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3095,"p95_latency_ms":3783,"p99_latency_ms":4058,"time_to_first_token_ms":576,"total_runtime_ms":3439,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0152,"confidence_interval":0.9352,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1885,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3092,"p95_latency_ms":3779,"p99_latency_ms":4054,"time_to_first_token_ms":586,"total_runtime_ms":3435,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.028,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7718,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3016,"p95_latency_ms":3686,"p99_latency_ms":3954,"time_to_first_token_ms":576,"total_runtime_ms":3351,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.022,"confidence_interval":0.9489,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0013,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3080,"p95_latency_ms":3765,"p99_latency_ms":4038,"time_to_first_token_ms":597,"total_runtime_ms":3422,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.026,"confidence_interval":0.938,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8772,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1509,"p95_latency_ms":1845,"p99_latency_ms":1979,"time_to_first_token_ms":536,"total_runtime_ms":1677,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0112,"confidence_interval":0.9479,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7229,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1513,"p95_latency_ms":1849,"p99_latency_ms":1984,"time_to_first_token_ms":528,"total_runtime_ms":1681,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0165,"confidence_interval":0.9408,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0512,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1590,"p95_latency_ms":1944,"p99_latency_ms":2085,"time_to_first_token_ms":530,"total_runtime_ms":1767,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0275,"confidence_interval":0.9406,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.884,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1518,"p95_latency_ms":1855,"p99_latency_ms":1990,"time_to_first_token_ms":544,"total_runtime_ms":1687,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0108,"confidence_interval":0.933,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7453,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1518,"p95_latency_ms":1855,"p99_latency_ms":1990,"time_to_first_token_ms":561,"total_runtime_ms":1686,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0279,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3781,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1532,"p95_latency_ms":1873,"p99_latency_ms":2009,"time_to_first_token_ms":535,"total_runtime_ms":1703,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.021,"confidence_interval":0.9325,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1915,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1587,"p95_latency_ms":1940,"p99_latency_ms":2081,"time_to_first_token_ms":555,"total_runtime_ms":1763,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0189,"confidence_interval":0.9377,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6503,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1561,"p95_latency_ms":1907,"p99_latency_ms":2046,"time_to_first_token_ms":534,"total_runtime_ms":1734,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0227,"confidence_interval":0.9487,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4444,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1549,"p95_latency_ms":1893,"p99_latency_ms":2031,"time_to_first_token_ms":531,"total_runtime_ms":1721,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0217,"confidence_interval":0.949,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.251,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|low|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1502,"p95_latency_ms":1836,"p99_latency_ms":1970,"time_to_first_token_ms":552,"total_runtime_ms":1669,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0199,"confidence_interval":0.9444,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1558,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":642,"p95_latency_ms":784,"p99_latency_ms":841,"time_to_first_token_ms":175,"total_runtime_ms":713,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0256,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2833,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":646,"p95_latency_ms":789,"p99_latency_ms":847,"time_to_first_token_ms":164,"total_runtime_ms":718,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0253,"confidence_interval":0.942,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8058,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":601,"p95_latency_ms":734,"p99_latency_ms":788,"time_to_first_token_ms":166,"total_runtime_ms":668,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0175,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3295,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":780,"time_to_first_token_ms":180,"total_runtime_ms":661,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0102,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3285,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":162,"total_runtime_ms":650,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0215,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5089,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":581,"p95_latency_ms":711,"p99_latency_ms":762,"time_to_first_token_ms":190,"total_runtime_ms":646,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0194,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2151,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":646,"p95_latency_ms":789,"p99_latency_ms":846,"time_to_first_token_ms":187,"total_runtime_ms":717,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0165,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8313,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":602,"p95_latency_ms":735,"p99_latency_ms":789,"time_to_first_token_ms":177,"total_runtime_ms":668,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0284,"confidence_interval":0.9454,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6306,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":657,"p95_latency_ms":803,"p99_latency_ms":861,"time_to_first_token_ms":174,"total_runtime_ms":730,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0236,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2599,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":633,"p95_latency_ms":774,"p99_latency_ms":830,"time_to_first_token_ms":173,"total_runtime_ms":703,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.02,"confidence_interval":0.9438,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0082,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":438,"p99_latency_ms":469,"time_to_first_token_ms":142,"total_runtime_ms":398,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0121,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8288,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":351,"p95_latency_ms":429,"p99_latency_ms":460,"time_to_first_token_ms":139,"total_runtime_ms":390,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0271,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2797,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":407,"time_to_first_token_ms":143,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0218,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0456,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":495,"time_to_first_token_ms":132,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0239,"confidence_interval":0.9407,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2794,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":356,"p95_latency_ms":435,"p99_latency_ms":467,"time_to_first_token_ms":136,"total_runtime_ms":396,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1938,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":334,"p95_latency_ms":408,"p99_latency_ms":438,"time_to_first_token_ms":141,"total_runtime_ms":371,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0278,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6457,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":353,"p95_latency_ms":431,"p99_latency_ms":463,"time_to_first_token_ms":126,"total_runtime_ms":392,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0185,"confidence_interval":0.9402,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1583,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":377,"p99_latency_ms":404,"time_to_first_token_ms":132,"total_runtime_ms":343,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0255,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2981,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":372,"p95_latency_ms":454,"p99_latency_ms":487,"time_to_first_token_ms":123,"total_runtime_ms":413,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0185,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2893,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":284,"p95_latency_ms":347,"p99_latency_ms":373,"time_to_first_token_ms":130,"total_runtime_ms":316,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0271,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7361,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":817,"p95_latency_ms":998,"p99_latency_ms":1071,"time_to_first_token_ms":220,"total_runtime_ms":908,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0255,"confidence_interval":0.934,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7741,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":215,"total_runtime_ms":849,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0238,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5616,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":839,"p95_latency_ms":1025,"p99_latency_ms":1100,"time_to_first_token_ms":199,"total_runtime_ms":932,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0277,"confidence_interval":0.9463,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3507,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":775,"p95_latency_ms":947,"p99_latency_ms":1016,"time_to_first_token_ms":216,"total_runtime_ms":861,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0113,"confidence_interval":0.9494,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5591,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":829,"p95_latency_ms":1013,"p99_latency_ms":1087,"time_to_first_token_ms":198,"total_runtime_ms":921,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0167,"confidence_interval":0.9326,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7439,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":797,"p95_latency_ms":974,"p99_latency_ms":1045,"time_to_first_token_ms":222,"total_runtime_ms":886,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.027,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7952,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":771,"p95_latency_ms":942,"p99_latency_ms":1011,"time_to_first_token_ms":201,"total_runtime_ms":857,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0231,"confidence_interval":0.9409,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4824,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":856,"p95_latency_ms":1046,"p99_latency_ms":1122,"time_to_first_token_ms":220,"total_runtime_ms":951,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0277,"confidence_interval":0.9489,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3262,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":860,"p95_latency_ms":1051,"p99_latency_ms":1127,"time_to_first_token_ms":197,"total_runtime_ms":955,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0185,"confidence_interval":0.9462,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6544,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1014,"p99_latency_ms":1088,"time_to_first_token_ms":216,"total_runtime_ms":922,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0182,"confidence_interval":0.9419,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7139,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":152,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9462,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9814,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":399,"p95_latency_ms":488,"p99_latency_ms":524,"time_to_first_token_ms":174,"total_runtime_ms":444,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0195,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2361,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":399,"p95_latency_ms":488,"p99_latency_ms":523,"time_to_first_token_ms":171,"total_runtime_ms":443,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.017,"confidence_interval":0.9355,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3476,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":430,"p95_latency_ms":526,"p99_latency_ms":564,"time_to_first_token_ms":173,"total_runtime_ms":478,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0122,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9263,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":399,"p95_latency_ms":488,"p99_latency_ms":524,"time_to_first_token_ms":178,"total_runtime_ms":444,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0196,"confidence_interval":0.9404,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.372,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":436,"p95_latency_ms":533,"p99_latency_ms":572,"time_to_first_token_ms":145,"total_runtime_ms":485,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0176,"confidence_interval":0.9436,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3787,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":416,"p95_latency_ms":509,"p99_latency_ms":546,"time_to_first_token_ms":152,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0103,"confidence_interval":0.9367,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9965,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":145,"total_runtime_ms":453,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0106,"confidence_interval":0.9435,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5248,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":468,"p95_latency_ms":572,"p99_latency_ms":614,"time_to_first_token_ms":173,"total_runtime_ms":520,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0145,"confidence_interval":0.9433,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.215,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":468,"p95_latency_ms":573,"p99_latency_ms":614,"time_to_first_token_ms":166,"total_runtime_ms":520,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0281,"confidence_interval":0.9362,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4046,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1623,"p95_latency_ms":1984,"p99_latency_ms":2128,"time_to_first_token_ms":348,"total_runtime_ms":1803,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0285,"confidence_interval":0.932,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1352,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1618,"p95_latency_ms":1978,"p99_latency_ms":2121,"time_to_first_token_ms":323,"total_runtime_ms":1798,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5071,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1523,"p95_latency_ms":1861,"p99_latency_ms":1996,"time_to_first_token_ms":320,"total_runtime_ms":1692,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0277,"confidence_interval":0.9483,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5942,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1592,"p95_latency_ms":1946,"p99_latency_ms":2087,"time_to_first_token_ms":326,"total_runtime_ms":1769,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.021,"confidence_interval":0.9371,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3608,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1562,"p95_latency_ms":1909,"p99_latency_ms":2048,"time_to_first_token_ms":343,"total_runtime_ms":1735,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0214,"confidence_interval":0.9315,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9123,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1552,"p95_latency_ms":1897,"p99_latency_ms":2035,"time_to_first_token_ms":328,"total_runtime_ms":1724,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0173,"confidence_interval":0.9307,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2545,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1528,"p95_latency_ms":1867,"p99_latency_ms":2003,"time_to_first_token_ms":342,"total_runtime_ms":1697,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0186,"confidence_interval":0.9399,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9706,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1561,"p95_latency_ms":1908,"p99_latency_ms":2047,"time_to_first_token_ms":326,"total_runtime_ms":1735,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0126,"confidence_interval":0.9326,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7905,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1557,"p95_latency_ms":1903,"p99_latency_ms":2042,"time_to_first_token_ms":323,"total_runtime_ms":1730,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0293,"confidence_interval":0.9382,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9586,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1600,"p95_latency_ms":1956,"p99_latency_ms":2098,"time_to_first_token_ms":346,"total_runtime_ms":1778,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0282,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9901,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":801,"p95_latency_ms":979,"p99_latency_ms":1050,"time_to_first_token_ms":283,"total_runtime_ms":890,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0274,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0198,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":854,"p95_latency_ms":1044,"p99_latency_ms":1120,"time_to_first_token_ms":276,"total_runtime_ms":949,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0295,"confidence_interval":0.9362,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9301,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":838,"p95_latency_ms":1024,"p99_latency_ms":1099,"time_to_first_token_ms":292,"total_runtime_ms":931,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.018,"confidence_interval":0.9481,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2052,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":774,"p95_latency_ms":946,"p99_latency_ms":1014,"time_to_first_token_ms":301,"total_runtime_ms":860,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0252,"confidence_interval":0.9472,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3368,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":846,"p95_latency_ms":1034,"p99_latency_ms":1109,"time_to_first_token_ms":289,"total_runtime_ms":940,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0288,"confidence_interval":0.9379,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5524,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":792,"p95_latency_ms":968,"p99_latency_ms":1038,"time_to_first_token_ms":305,"total_runtime_ms":880,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.021,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3624,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":840,"p95_latency_ms":1026,"p99_latency_ms":1101,"time_to_first_token_ms":276,"total_runtime_ms":933,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0264,"confidence_interval":0.9347,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5945,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":802,"p95_latency_ms":980,"p99_latency_ms":1051,"time_to_first_token_ms":281,"total_runtime_ms":891,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0277,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9169,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":768,"p95_latency_ms":939,"p99_latency_ms":1008,"time_to_first_token_ms":276,"total_runtime_ms":854,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.011,"confidence_interval":0.9475,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7864,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":773,"p95_latency_ms":945,"p99_latency_ms":1014,"time_to_first_token_ms":303,"total_runtime_ms":859,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0147,"confidence_interval":0.9328,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3334,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2328,"p95_latency_ms":2845,"p99_latency_ms":3052,"time_to_first_token_ms":469,"total_runtime_ms":2586,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0269,"confidence_interval":0.937,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6141,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2293,"p95_latency_ms":2803,"p99_latency_ms":3007,"time_to_first_token_ms":475,"total_runtime_ms":2548,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.024,"confidence_interval":0.9366,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2455,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2293,"p95_latency_ms":2803,"p99_latency_ms":3006,"time_to_first_token_ms":475,"total_runtime_ms":2548,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0216,"confidence_interval":0.9368,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3621,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2282,"p95_latency_ms":2789,"p99_latency_ms":2992,"time_to_first_token_ms":460,"total_runtime_ms":2536,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0262,"confidence_interval":0.9471,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6487,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2253,"p95_latency_ms":2754,"p99_latency_ms":2954,"time_to_first_token_ms":463,"total_runtime_ms":2503,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0215,"confidence_interval":0.9463,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6491,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2250,"p95_latency_ms":2750,"p99_latency_ms":2950,"time_to_first_token_ms":471,"total_runtime_ms":2500,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9358,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3472,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2291,"p95_latency_ms":2800,"p99_latency_ms":3004,"time_to_first_token_ms":464,"total_runtime_ms":2546,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.026,"confidence_interval":0.9463,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0168,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2257,"p95_latency_ms":2758,"p99_latency_ms":2959,"time_to_first_token_ms":479,"total_runtime_ms":2508,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6504,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2276,"p95_latency_ms":2782,"p99_latency_ms":2984,"time_to_first_token_ms":478,"total_runtime_ms":2529,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0247,"confidence_interval":0.9351,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.884,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2296,"p95_latency_ms":2806,"p99_latency_ms":3011,"time_to_first_token_ms":454,"total_runtime_ms":2551,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0214,"confidence_interval":0.9485,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9201,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1140,"p95_latency_ms":1393,"p99_latency_ms":1495,"time_to_first_token_ms":406,"total_runtime_ms":1267,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.9492,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.102,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1168,"p95_latency_ms":1427,"p99_latency_ms":1531,"time_to_first_token_ms":401,"total_runtime_ms":1298,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0119,"confidence_interval":0.9346,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2894,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1201,"p95_latency_ms":1468,"p99_latency_ms":1575,"time_to_first_token_ms":428,"total_runtime_ms":1335,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0205,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3361,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1186,"p95_latency_ms":1450,"p99_latency_ms":1555,"time_to_first_token_ms":415,"total_runtime_ms":1318,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0229,"confidence_interval":0.9322,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.96,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1221,"p95_latency_ms":1493,"p99_latency_ms":1601,"time_to_first_token_ms":401,"total_runtime_ms":1357,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0123,"confidence_interval":0.9338,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9948,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1146,"p95_latency_ms":1401,"p99_latency_ms":1503,"time_to_first_token_ms":427,"total_runtime_ms":1273,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0126,"confidence_interval":0.9438,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4326,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1174,"p95_latency_ms":1435,"p99_latency_ms":1539,"time_to_first_token_ms":432,"total_runtime_ms":1304,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0151,"confidence_interval":0.9492,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.321,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1221,"p95_latency_ms":1493,"p99_latency_ms":1601,"time_to_first_token_ms":410,"total_runtime_ms":1357,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0122,"confidence_interval":0.9424,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.032,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1138,"p95_latency_ms":1391,"p99_latency_ms":1492,"time_to_first_token_ms":429,"total_runtime_ms":1264,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0281,"confidence_interval":0.9393,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8247,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1161,"p95_latency_ms":1420,"p99_latency_ms":1523,"time_to_first_token_ms":433,"total_runtime_ms":1290,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0123,"confidence_interval":0.9412,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3798,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":493,"p95_latency_ms":602,"p99_latency_ms":646,"time_to_first_token_ms":140,"total_runtime_ms":547,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4129,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":446,"p95_latency_ms":546,"p99_latency_ms":585,"time_to_first_token_ms":137,"total_runtime_ms":496,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.017,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5431,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":474,"p95_latency_ms":579,"p99_latency_ms":621,"time_to_first_token_ms":163,"total_runtime_ms":526,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.016,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4496,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":474,"p95_latency_ms":579,"p99_latency_ms":621,"time_to_first_token_ms":151,"total_runtime_ms":527,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0266,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2274,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":447,"p95_latency_ms":547,"p99_latency_ms":586,"time_to_first_token_ms":135,"total_runtime_ms":497,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0276,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0726,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":521,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":153,"total_runtime_ms":578,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0227,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2101,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":155,"total_runtime_ms":488,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0216,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9587,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":433,"p95_latency_ms":529,"p99_latency_ms":567,"time_to_first_token_ms":135,"total_runtime_ms":481,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0126,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0483,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":438,"p95_latency_ms":535,"p99_latency_ms":574,"time_to_first_token_ms":139,"total_runtime_ms":486,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0229,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9638,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":470,"p95_latency_ms":575,"p99_latency_ms":616,"time_to_first_token_ms":137,"total_runtime_ms":522,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0172,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1421,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":288,"p95_latency_ms":352,"p99_latency_ms":378,"time_to_first_token_ms":102,"total_runtime_ms":320,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0224,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8982,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":230,"p95_latency_ms":282,"p99_latency_ms":302,"time_to_first_token_ms":89,"total_runtime_ms":256,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0181,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0701,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":318,"p95_latency_ms":389,"p99_latency_ms":417,"time_to_first_token_ms":110,"total_runtime_ms":353,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0297,"confidence_interval":0.9419,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3939,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":303,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":92,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0272,"confidence_interval":0.9347,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4441,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":239,"p95_latency_ms":292,"p99_latency_ms":314,"time_to_first_token_ms":117,"total_runtime_ms":266,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0292,"confidence_interval":0.9368,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1597,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":225,"p95_latency_ms":275,"p99_latency_ms":295,"time_to_first_token_ms":101,"total_runtime_ms":250,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0206,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5974,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":264,"p95_latency_ms":322,"p99_latency_ms":346,"time_to_first_token_ms":112,"total_runtime_ms":293,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.019,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.246,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":234,"p95_latency_ms":287,"p99_latency_ms":307,"time_to_first_token_ms":106,"total_runtime_ms":261,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0217,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9971,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":243,"p95_latency_ms":297,"p99_latency_ms":318,"time_to_first_token_ms":94,"total_runtime_ms":270,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0235,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6954,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":268,"p95_latency_ms":327,"p99_latency_ms":351,"time_to_first_token_ms":120,"total_runtime_ms":298,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0112,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4279,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":592,"p95_latency_ms":724,"p99_latency_ms":777,"time_to_first_token_ms":163,"total_runtime_ms":658,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0221,"confidence_interval":0.934,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8001,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":628,"p95_latency_ms":768,"p99_latency_ms":824,"time_to_first_token_ms":160,"total_runtime_ms":698,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0282,"confidence_interval":0.948,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2466,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":673,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":180,"total_runtime_ms":747,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0284,"confidence_interval":0.9458,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.305,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":593,"p95_latency_ms":724,"p99_latency_ms":777,"time_to_first_token_ms":191,"total_runtime_ms":659,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0211,"confidence_interval":0.9308,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7742,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":626,"p95_latency_ms":765,"p99_latency_ms":821,"time_to_first_token_ms":183,"total_runtime_ms":696,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0236,"confidence_interval":0.9306,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.206,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":587,"p95_latency_ms":718,"p99_latency_ms":770,"time_to_first_token_ms":159,"total_runtime_ms":653,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0156,"confidence_interval":0.9431,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9299,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":662,"p95_latency_ms":809,"p99_latency_ms":867,"time_to_first_token_ms":192,"total_runtime_ms":735,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0177,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7555,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":615,"p95_latency_ms":751,"p99_latency_ms":806,"time_to_first_token_ms":169,"total_runtime_ms":683,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0111,"confidence_interval":0.9328,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0057,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":651,"p95_latency_ms":795,"p99_latency_ms":853,"time_to_first_token_ms":179,"total_runtime_ms":723,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0191,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5336,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":630,"p95_latency_ms":770,"p99_latency_ms":826,"time_to_first_token_ms":162,"total_runtime_ms":700,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0253,"confidence_interval":0.941,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6963,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":343,"p95_latency_ms":419,"p99_latency_ms":450,"time_to_first_token_ms":124,"total_runtime_ms":381,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0139,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3273,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":338,"p95_latency_ms":413,"p99_latency_ms":443,"time_to_first_token_ms":132,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0249,"confidence_interval":0.9352,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8276,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":361,"p95_latency_ms":441,"p99_latency_ms":474,"time_to_first_token_ms":125,"total_runtime_ms":401,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0134,"confidence_interval":0.9452,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7119,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":483,"time_to_first_token_ms":147,"total_runtime_ms":409,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0224,"confidence_interval":0.9335,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7949,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":335,"p95_latency_ms":410,"p99_latency_ms":440,"time_to_first_token_ms":145,"total_runtime_ms":373,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0295,"confidence_interval":0.9346,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4175,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":471,"time_to_first_token_ms":122,"total_runtime_ms":399,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0252,"confidence_interval":0.947,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7124,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":332,"p95_latency_ms":406,"p99_latency_ms":436,"time_to_first_token_ms":146,"total_runtime_ms":369,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0159,"confidence_interval":0.9499,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2275,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":424,"time_to_first_token_ms":128,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0177,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5478,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":348,"p95_latency_ms":425,"p99_latency_ms":456,"time_to_first_token_ms":131,"total_runtime_ms":386,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.013,"confidence_interval":0.9483,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5995,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":319,"p95_latency_ms":390,"p99_latency_ms":418,"time_to_first_token_ms":122,"total_runtime_ms":354,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0138,"confidence_interval":0.9315,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1743,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1167,"p95_latency_ms":1427,"p99_latency_ms":1531,"time_to_first_token_ms":256,"total_runtime_ms":1297,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0253,"confidence_interval":0.9352,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0321,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1244,"p95_latency_ms":1520,"p99_latency_ms":1631,"time_to_first_token_ms":282,"total_runtime_ms":1382,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0141,"confidence_interval":0.9418,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6719,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1180,"p95_latency_ms":1442,"p99_latency_ms":1547,"time_to_first_token_ms":283,"total_runtime_ms":1311,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0148,"confidence_interval":0.938,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.11,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1214,"p95_latency_ms":1484,"p99_latency_ms":1592,"time_to_first_token_ms":278,"total_runtime_ms":1349,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0285,"confidence_interval":0.9388,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3633,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1183,"p95_latency_ms":1446,"p99_latency_ms":1551,"time_to_first_token_ms":258,"total_runtime_ms":1314,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0124,"confidence_interval":0.9482,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7994,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1241,"p95_latency_ms":1517,"p99_latency_ms":1628,"time_to_first_token_ms":268,"total_runtime_ms":1379,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0155,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4698,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1172,"p95_latency_ms":1432,"p99_latency_ms":1537,"time_to_first_token_ms":263,"total_runtime_ms":1302,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0274,"confidence_interval":0.936,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.634,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1194,"p95_latency_ms":1460,"p99_latency_ms":1566,"time_to_first_token_ms":274,"total_runtime_ms":1327,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0178,"confidence_interval":0.9498,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.318,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1168,"p95_latency_ms":1428,"p99_latency_ms":1532,"time_to_first_token_ms":263,"total_runtime_ms":1298,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.934,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.721,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1186,"p95_latency_ms":1450,"p99_latency_ms":1555,"time_to_first_token_ms":258,"total_runtime_ms":1318,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0117,"confidence_interval":0.9317,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4363,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":643,"p95_latency_ms":785,"p99_latency_ms":843,"time_to_first_token_ms":238,"total_runtime_ms":714,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0201,"confidence_interval":0.9382,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9402,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":844,"time_to_first_token_ms":227,"total_runtime_ms":715,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0147,"confidence_interval":0.9312,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4748,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":652,"p95_latency_ms":797,"p99_latency_ms":855,"time_to_first_token_ms":243,"total_runtime_ms":724,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0178,"confidence_interval":0.9393,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6018,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":669,"p95_latency_ms":817,"p99_latency_ms":877,"time_to_first_token_ms":221,"total_runtime_ms":743,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0246,"confidence_interval":0.9321,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5439,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":672,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":220,"total_runtime_ms":747,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0296,"confidence_interval":0.9463,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4961,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":614,"p95_latency_ms":751,"p99_latency_ms":805,"time_to_first_token_ms":220,"total_runtime_ms":682,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0224,"confidence_interval":0.9496,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0476,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":654,"p95_latency_ms":799,"p99_latency_ms":857,"time_to_first_token_ms":213,"total_runtime_ms":727,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0222,"confidence_interval":0.941,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4977,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":595,"p95_latency_ms":728,"p99_latency_ms":780,"time_to_first_token_ms":212,"total_runtime_ms":661,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0116,"confidence_interval":0.9437,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4779,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":875,"time_to_first_token_ms":228,"total_runtime_ms":741,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.024,"confidence_interval":0.9493,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9875,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":600,"p95_latency_ms":734,"p99_latency_ms":787,"time_to_first_token_ms":215,"total_runtime_ms":667,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0109,"confidence_interval":0.9472,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5344,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1754,"p95_latency_ms":2143,"p99_latency_ms":2299,"time_to_first_token_ms":363,"total_runtime_ms":1949,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0264,"confidence_interval":0.9451,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6695,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1778,"p95_latency_ms":2174,"p99_latency_ms":2332,"time_to_first_token_ms":377,"total_runtime_ms":1976,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0259,"confidence_interval":0.9446,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8574,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1722,"p95_latency_ms":2105,"p99_latency_ms":2258,"time_to_first_token_ms":381,"total_runtime_ms":1914,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0203,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7209,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1713,"p95_latency_ms":2093,"p99_latency_ms":2245,"time_to_first_token_ms":375,"total_runtime_ms":1903,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0296,"confidence_interval":0.934,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6184,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1779,"p95_latency_ms":2175,"p99_latency_ms":2333,"time_to_first_token_ms":378,"total_runtime_ms":1977,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0259,"confidence_interval":0.9429,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2979,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1746,"p95_latency_ms":2133,"p99_latency_ms":2289,"time_to_first_token_ms":376,"total_runtime_ms":1939,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0274,"confidence_interval":0.949,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8279,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1771,"p95_latency_ms":2164,"p99_latency_ms":2322,"time_to_first_token_ms":371,"total_runtime_ms":1968,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0113,"confidence_interval":0.944,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1217,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1719,"p95_latency_ms":2100,"p99_latency_ms":2253,"time_to_first_token_ms":350,"total_runtime_ms":1909,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0207,"confidence_interval":0.9348,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9137,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1752,"p95_latency_ms":2142,"p99_latency_ms":2298,"time_to_first_token_ms":378,"total_runtime_ms":1947,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0113,"confidence_interval":0.9323,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.662,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1694,"p95_latency_ms":2070,"p99_latency_ms":2221,"time_to_first_token_ms":351,"total_runtime_ms":1882,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0158,"confidence_interval":0.9388,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6807,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":927,"p95_latency_ms":1133,"p99_latency_ms":1215,"time_to_first_token_ms":312,"total_runtime_ms":1030,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.9317,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8561,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":847,"p95_latency_ms":1035,"p99_latency_ms":1111,"time_to_first_token_ms":319,"total_runtime_ms":941,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0188,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0544,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":896,"p95_latency_ms":1096,"p99_latency_ms":1175,"time_to_first_token_ms":307,"total_runtime_ms":996,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.029,"confidence_interval":0.932,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1138,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":892,"p95_latency_ms":1090,"p99_latency_ms":1170,"time_to_first_token_ms":317,"total_runtime_ms":991,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0291,"confidence_interval":0.9331,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6718,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":934,"p95_latency_ms":1142,"p99_latency_ms":1225,"time_to_first_token_ms":322,"total_runtime_ms":1038,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0259,"confidence_interval":0.9366,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2238,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":857,"p95_latency_ms":1047,"p99_latency_ms":1123,"time_to_first_token_ms":325,"total_runtime_ms":952,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0173,"confidence_interval":0.9406,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0651,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":910,"p95_latency_ms":1112,"p99_latency_ms":1192,"time_to_first_token_ms":323,"total_runtime_ms":1011,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0167,"confidence_interval":0.9473,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3881,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":886,"p95_latency_ms":1083,"p99_latency_ms":1162,"time_to_first_token_ms":317,"total_runtime_ms":985,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0181,"confidence_interval":0.9318,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4188,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":858,"p95_latency_ms":1048,"p99_latency_ms":1124,"time_to_first_token_ms":310,"total_runtime_ms":953,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0285,"confidence_interval":0.9325,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4186,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":937,"p95_latency_ms":1145,"p99_latency_ms":1228,"time_to_first_token_ms":338,"total_runtime_ms":1041,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9446,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9621,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":412,"p95_latency_ms":504,"p99_latency_ms":540,"time_to_first_token_ms":122,"total_runtime_ms":458,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0155,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7119,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":398,"p95_latency_ms":487,"p99_latency_ms":522,"time_to_first_token_ms":134,"total_runtime_ms":443,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0232,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3702,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":479,"time_to_first_token_ms":127,"total_runtime_ms":406,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0227,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3191,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":513,"time_to_first_token_ms":129,"total_runtime_ms":434,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0212,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0967,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":127,"total_runtime_ms":435,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0123,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6817,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":151,"total_runtime_ms":475,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1346,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":438,"p95_latency_ms":536,"p99_latency_ms":575,"time_to_first_token_ms":136,"total_runtime_ms":487,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0286,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0934,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":508,"time_to_first_token_ms":137,"total_runtime_ms":430,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0143,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2165,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":432,"p95_latency_ms":528,"p99_latency_ms":566,"time_to_first_token_ms":153,"total_runtime_ms":480,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0131,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5452,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":396,"p95_latency_ms":485,"p99_latency_ms":520,"time_to_first_token_ms":152,"total_runtime_ms":440,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0228,"confidence_interval":0.9448,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2644,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":210,"p95_latency_ms":257,"p99_latency_ms":275,"time_to_first_token_ms":95,"total_runtime_ms":233,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0118,"confidence_interval":0.9466,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0805,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":272,"p95_latency_ms":333,"p99_latency_ms":357,"time_to_first_token_ms":105,"total_runtime_ms":303,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0282,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1079,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":179,"p95_latency_ms":219,"p99_latency_ms":235,"time_to_first_token_ms":110,"total_runtime_ms":199,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0228,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6938,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":211,"p95_latency_ms":258,"p99_latency_ms":277,"time_to_first_token_ms":88,"total_runtime_ms":235,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0162,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1978,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":191,"p95_latency_ms":233,"p99_latency_ms":250,"time_to_first_token_ms":88,"total_runtime_ms":212,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0206,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0713,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":225,"p95_latency_ms":275,"p99_latency_ms":295,"time_to_first_token_ms":107,"total_runtime_ms":250,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0221,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0704,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":270,"p95_latency_ms":330,"p99_latency_ms":354,"time_to_first_token_ms":100,"total_runtime_ms":300,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0239,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1666,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":210,"p95_latency_ms":256,"p99_latency_ms":275,"time_to_first_token_ms":91,"total_runtime_ms":233,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0263,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1871,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":220,"p95_latency_ms":268,"p99_latency_ms":288,"time_to_first_token_ms":93,"total_runtime_ms":244,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8236,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":281,"p95_latency_ms":344,"p99_latency_ms":369,"time_to_first_token_ms":91,"total_runtime_ms":313,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0142,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6407,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":562,"p95_latency_ms":687,"p99_latency_ms":737,"time_to_first_token_ms":152,"total_runtime_ms":625,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0272,"confidence_interval":0.9471,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7938,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":516,"p95_latency_ms":630,"p99_latency_ms":676,"time_to_first_token_ms":154,"total_runtime_ms":573,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0148,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7117,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":534,"p95_latency_ms":653,"p99_latency_ms":700,"time_to_first_token_ms":165,"total_runtime_ms":594,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0217,"confidence_interval":0.9369,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5721,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":561,"p95_latency_ms":686,"p99_latency_ms":736,"time_to_first_token_ms":155,"total_runtime_ms":623,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0102,"confidence_interval":0.9344,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7274,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":160,"total_runtime_ms":584,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0291,"confidence_interval":0.9386,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8932,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":558,"p95_latency_ms":682,"p99_latency_ms":731,"time_to_first_token_ms":169,"total_runtime_ms":620,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0127,"confidence_interval":0.9446,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9449,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":555,"p95_latency_ms":678,"p99_latency_ms":727,"time_to_first_token_ms":172,"total_runtime_ms":616,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0257,"confidence_interval":0.9466,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5994,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":554,"p95_latency_ms":677,"p99_latency_ms":726,"time_to_first_token_ms":154,"total_runtime_ms":615,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0155,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3724,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":672,"time_to_first_token_ms":143,"total_runtime_ms":569,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5225,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":543,"p95_latency_ms":664,"p99_latency_ms":712,"time_to_first_token_ms":163,"total_runtime_ms":603,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0212,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0513,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":291,"p99_latency_ms":312,"time_to_first_token_ms":110,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0194,"confidence_interval":0.9375,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8673,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":313,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":103,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0287,"confidence_interval":0.9434,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7129,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":242,"p95_latency_ms":296,"p99_latency_ms":317,"time_to_first_token_ms":108,"total_runtime_ms":269,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0206,"confidence_interval":0.9433,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9556,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":242,"p95_latency_ms":296,"p99_latency_ms":318,"time_to_first_token_ms":108,"total_runtime_ms":269,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0272,"confidence_interval":0.9497,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8695,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":313,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":110,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0296,"confidence_interval":0.9431,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8181,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":327,"p95_latency_ms":400,"p99_latency_ms":429,"time_to_first_token_ms":110,"total_runtime_ms":364,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0223,"confidence_interval":0.9412,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7344,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":270,"p95_latency_ms":330,"p99_latency_ms":354,"time_to_first_token_ms":128,"total_runtime_ms":300,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.03,"confidence_interval":0.9319,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8407,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":303,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":118,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0145,"confidence_interval":0.945,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5591,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":278,"p95_latency_ms":340,"p99_latency_ms":364,"time_to_first_token_ms":124,"total_runtime_ms":309,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0199,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3102,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":302,"p95_latency_ms":369,"p99_latency_ms":396,"time_to_first_token_ms":114,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0155,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7115,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1041,"p95_latency_ms":1272,"p99_latency_ms":1365,"time_to_first_token_ms":235,"total_runtime_ms":1156,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0165,"confidence_interval":0.9486,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2743,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1001,"p95_latency_ms":1224,"p99_latency_ms":1313,"time_to_first_token_ms":256,"total_runtime_ms":1113,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0112,"confidence_interval":0.9331,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6623,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1025,"p95_latency_ms":1252,"p99_latency_ms":1344,"time_to_first_token_ms":235,"total_runtime_ms":1139,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0123,"confidence_interval":0.934,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0045,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":989,"p95_latency_ms":1209,"p99_latency_ms":1297,"time_to_first_token_ms":245,"total_runtime_ms":1099,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0146,"confidence_interval":0.943,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7348,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":992,"p95_latency_ms":1213,"p99_latency_ms":1301,"time_to_first_token_ms":229,"total_runtime_ms":1103,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9479,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0414,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1002,"p95_latency_ms":1224,"p99_latency_ms":1313,"time_to_first_token_ms":233,"total_runtime_ms":1113,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0116,"confidence_interval":0.9429,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4012,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1022,"p95_latency_ms":1250,"p99_latency_ms":1340,"time_to_first_token_ms":253,"total_runtime_ms":1136,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0283,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.557,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1027,"p95_latency_ms":1255,"p99_latency_ms":1346,"time_to_first_token_ms":223,"total_runtime_ms":1141,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0172,"confidence_interval":0.9429,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1821,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1023,"p95_latency_ms":1250,"p99_latency_ms":1341,"time_to_first_token_ms":238,"total_runtime_ms":1136,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0221,"confidence_interval":0.9308,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2108,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":981,"p95_latency_ms":1199,"p99_latency_ms":1286,"time_to_first_token_ms":241,"total_runtime_ms":1090,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0106,"confidence_interval":0.932,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7061,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":755,"time_to_first_token_ms":200,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0223,"confidence_interval":0.9357,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2034,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":189,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0239,"confidence_interval":0.9446,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6624,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":743,"time_to_first_token_ms":195,"total_runtime_ms":630,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0207,"confidence_interval":0.9415,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0124,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":573,"p95_latency_ms":701,"p99_latency_ms":752,"time_to_first_token_ms":191,"total_runtime_ms":637,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9498,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.271,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":484,"p95_latency_ms":592,"p99_latency_ms":635,"time_to_first_token_ms":192,"total_runtime_ms":538,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0102,"confidence_interval":0.9331,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.3142,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":506,"p95_latency_ms":618,"p99_latency_ms":663,"time_to_first_token_ms":189,"total_runtime_ms":562,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0282,"confidence_interval":0.9391,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9566,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":493,"p95_latency_ms":603,"p99_latency_ms":647,"time_to_first_token_ms":204,"total_runtime_ms":548,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0231,"confidence_interval":0.9324,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5506,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":521,"p95_latency_ms":637,"p99_latency_ms":683,"time_to_first_token_ms":210,"total_runtime_ms":579,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0154,"confidence_interval":0.9421,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8158,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":566,"p95_latency_ms":691,"p99_latency_ms":742,"time_to_first_token_ms":203,"total_runtime_ms":628,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0178,"confidence_interval":0.9355,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.02,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":581,"p95_latency_ms":710,"p99_latency_ms":762,"time_to_first_token_ms":177,"total_runtime_ms":645,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9353,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6804,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1445,"p95_latency_ms":1766,"p99_latency_ms":1895,"time_to_first_token_ms":312,"total_runtime_ms":1606,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0153,"confidence_interval":0.945,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5726,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1475,"p95_latency_ms":1803,"p99_latency_ms":1934,"time_to_first_token_ms":324,"total_runtime_ms":1639,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0255,"confidence_interval":0.9331,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2069,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1496,"p95_latency_ms":1829,"p99_latency_ms":1962,"time_to_first_token_ms":329,"total_runtime_ms":1663,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.024,"confidence_interval":0.9376,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7241,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1497,"p95_latency_ms":1830,"p99_latency_ms":1963,"time_to_first_token_ms":305,"total_runtime_ms":1663,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0107,"confidence_interval":0.9331,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.516,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1497,"p95_latency_ms":1830,"p99_latency_ms":1963,"time_to_first_token_ms":308,"total_runtime_ms":1664,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.016,"confidence_interval":0.9432,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7858,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1465,"p95_latency_ms":1791,"p99_latency_ms":1921,"time_to_first_token_ms":321,"total_runtime_ms":1628,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0118,"confidence_interval":0.9383,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3473,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1431,"p95_latency_ms":1749,"p99_latency_ms":1877,"time_to_first_token_ms":313,"total_runtime_ms":1590,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0268,"confidence_interval":0.9446,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3672,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1512,"p95_latency_ms":1848,"p99_latency_ms":1982,"time_to_first_token_ms":306,"total_runtime_ms":1680,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0116,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8987,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1467,"p95_latency_ms":1793,"p99_latency_ms":1923,"time_to_first_token_ms":328,"total_runtime_ms":1630,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0144,"confidence_interval":0.9344,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.203,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1425,"p95_latency_ms":1742,"p99_latency_ms":1869,"time_to_first_token_ms":304,"total_runtime_ms":1583,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0275,"confidence_interval":0.9319,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0366,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":733,"p95_latency_ms":895,"p99_latency_ms":961,"time_to_first_token_ms":256,"total_runtime_ms":814,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0224,"confidence_interval":0.943,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8096,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":771,"p95_latency_ms":943,"p99_latency_ms":1012,"time_to_first_token_ms":267,"total_runtime_ms":857,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0204,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0076,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":794,"p95_latency_ms":970,"p99_latency_ms":1041,"time_to_first_token_ms":259,"total_runtime_ms":882,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0157,"confidence_interval":0.9401,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8242,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":286,"total_runtime_ms":895,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0109,"confidence_interval":0.9461,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.899,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":809,"p95_latency_ms":988,"p99_latency_ms":1060,"time_to_first_token_ms":263,"total_runtime_ms":899,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0268,"confidence_interval":0.9363,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1088,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":766,"p95_latency_ms":936,"p99_latency_ms":1004,"time_to_first_token_ms":255,"total_runtime_ms":851,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9359,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2774,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":748,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":259,"total_runtime_ms":831,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.017,"confidence_interval":0.9331,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3744,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":733,"p95_latency_ms":895,"p99_latency_ms":961,"time_to_first_token_ms":273,"total_runtime_ms":814,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.9336,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9439,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":722,"p95_latency_ms":882,"p99_latency_ms":946,"time_to_first_token_ms":263,"total_runtime_ms":802,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0159,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3538,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_minimal|high|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":743,"p95_latency_ms":908,"p99_latency_ms":974,"time_to_first_token_ms":264,"total_runtime_ms":825,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0173,"confidence_interval":0.9344,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5603,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":829,"p95_latency_ms":1013,"p99_latency_ms":1087,"time_to_first_token_ms":218,"total_runtime_ms":921,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0126,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0476,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":825,"p95_latency_ms":1008,"p99_latency_ms":1082,"time_to_first_token_ms":216,"total_runtime_ms":917,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0145,"confidence_interval":0.9348,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2498,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":914,"p95_latency_ms":1117,"p99_latency_ms":1198,"time_to_first_token_ms":215,"total_runtime_ms":1015,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0185,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4345,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":932,"p95_latency_ms":1139,"p99_latency_ms":1221,"time_to_first_token_ms":208,"total_runtime_ms":1035,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0293,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6943,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":915,"p95_latency_ms":1118,"p99_latency_ms":1200,"time_to_first_token_ms":197,"total_runtime_ms":1017,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0266,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1849,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":869,"p95_latency_ms":1062,"p99_latency_ms":1140,"time_to_first_token_ms":216,"total_runtime_ms":966,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0286,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7941,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":926,"p95_latency_ms":1131,"p99_latency_ms":1214,"time_to_first_token_ms":203,"total_runtime_ms":1028,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0199,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0819,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":920,"p95_latency_ms":1125,"p99_latency_ms":1206,"time_to_first_token_ms":206,"total_runtime_ms":1022,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0238,"confidence_interval":0.9321,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4266,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":914,"p95_latency_ms":1117,"p99_latency_ms":1198,"time_to_first_token_ms":219,"total_runtime_ms":1016,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0112,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3215,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":879,"p95_latency_ms":1075,"p99_latency_ms":1153,"time_to_first_token_ms":208,"total_runtime_ms":977,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.03,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3874,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":453,"p95_latency_ms":553,"p99_latency_ms":594,"time_to_first_token_ms":158,"total_runtime_ms":503,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0216,"confidence_interval":0.9346,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2938,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":146,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0152,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8452,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":520,"p95_latency_ms":635,"p99_latency_ms":682,"time_to_first_token_ms":151,"total_runtime_ms":578,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.013,"confidence_interval":0.9423,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3405,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":455,"p95_latency_ms":556,"p99_latency_ms":596,"time_to_first_token_ms":174,"total_runtime_ms":505,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0236,"confidence_interval":0.9416,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.06,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":504,"p95_latency_ms":617,"p99_latency_ms":661,"time_to_first_token_ms":150,"total_runtime_ms":560,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3027,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":453,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":175,"total_runtime_ms":503,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0274,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6005,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":543,"p95_latency_ms":664,"p99_latency_ms":712,"time_to_first_token_ms":164,"total_runtime_ms":604,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0159,"confidence_interval":0.9361,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4113,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":540,"p95_latency_ms":660,"p99_latency_ms":708,"time_to_first_token_ms":177,"total_runtime_ms":600,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.026,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1758,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":450,"p95_latency_ms":550,"p99_latency_ms":590,"time_to_first_token_ms":162,"total_runtime_ms":500,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0194,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5996,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":551,"p95_latency_ms":673,"p99_latency_ms":722,"time_to_first_token_ms":166,"total_runtime_ms":612,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9912,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1217,"p95_latency_ms":1487,"p99_latency_ms":1595,"time_to_first_token_ms":247,"total_runtime_ms":1352,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.026,"confidence_interval":0.9443,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0751,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1137,"p95_latency_ms":1389,"p99_latency_ms":1490,"time_to_first_token_ms":235,"total_runtime_ms":1263,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0198,"confidence_interval":0.933,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6909,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1147,"p95_latency_ms":1402,"p99_latency_ms":1504,"time_to_first_token_ms":234,"total_runtime_ms":1274,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0201,"confidence_interval":0.9335,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4513,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1167,"p95_latency_ms":1426,"p99_latency_ms":1530,"time_to_first_token_ms":259,"total_runtime_ms":1296,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0279,"confidence_interval":0.9475,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4773,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1138,"p95_latency_ms":1390,"p99_latency_ms":1492,"time_to_first_token_ms":263,"total_runtime_ms":1264,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0192,"confidence_interval":0.9368,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9653,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1181,"p95_latency_ms":1443,"p99_latency_ms":1548,"time_to_first_token_ms":234,"total_runtime_ms":1312,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0171,"confidence_interval":0.9333,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8946,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1210,"p95_latency_ms":1479,"p99_latency_ms":1587,"time_to_first_token_ms":259,"total_runtime_ms":1345,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0102,"confidence_interval":0.9334,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4479,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1212,"p95_latency_ms":1481,"p99_latency_ms":1588,"time_to_first_token_ms":265,"total_runtime_ms":1346,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0102,"confidence_interval":0.9371,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8675,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1182,"p95_latency_ms":1444,"p99_latency_ms":1549,"time_to_first_token_ms":264,"total_runtime_ms":1313,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0236,"confidence_interval":0.9332,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6126,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1117,"p95_latency_ms":1366,"p99_latency_ms":1465,"time_to_first_token_ms":239,"total_runtime_ms":1242,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0149,"confidence_interval":0.9328,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5331,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":636,"p95_latency_ms":778,"p99_latency_ms":834,"time_to_first_token_ms":196,"total_runtime_ms":707,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0181,"confidence_interval":0.937,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1531,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":630,"p95_latency_ms":770,"p99_latency_ms":826,"time_to_first_token_ms":196,"total_runtime_ms":700,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0109,"confidence_interval":0.9418,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4186,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":630,"p95_latency_ms":769,"p99_latency_ms":825,"time_to_first_token_ms":216,"total_runtime_ms":699,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0288,"confidence_interval":0.939,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2696,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":622,"p95_latency_ms":760,"p99_latency_ms":815,"time_to_first_token_ms":221,"total_runtime_ms":691,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0219,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7755,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":639,"p95_latency_ms":782,"p99_latency_ms":838,"time_to_first_token_ms":197,"total_runtime_ms":710,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0159,"confidence_interval":0.945,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3793,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":702,"p95_latency_ms":857,"p99_latency_ms":920,"time_to_first_token_ms":211,"total_runtime_ms":779,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0248,"confidence_interval":0.9323,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3419,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":843,"time_to_first_token_ms":190,"total_runtime_ms":715,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0229,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7805,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":651,"p95_latency_ms":796,"p99_latency_ms":854,"time_to_first_token_ms":213,"total_runtime_ms":724,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9418,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.092,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":696,"p95_latency_ms":851,"p99_latency_ms":913,"time_to_first_token_ms":219,"total_runtime_ms":773,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0202,"confidence_interval":0.9489,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0514,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":683,"p95_latency_ms":835,"p99_latency_ms":896,"time_to_first_token_ms":196,"total_runtime_ms":759,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0273,"confidence_interval":0.9478,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2146,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2299,"p95_latency_ms":2810,"p99_latency_ms":3014,"time_to_first_token_ms":408,"total_runtime_ms":2555,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0201,"confidence_interval":0.9445,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2785,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2326,"p95_latency_ms":2842,"p99_latency_ms":3049,"time_to_first_token_ms":436,"total_runtime_ms":2584,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0185,"confidence_interval":0.9305,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3327,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2303,"p95_latency_ms":2815,"p99_latency_ms":3020,"time_to_first_token_ms":424,"total_runtime_ms":2559,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0259,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2996,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2291,"p95_latency_ms":2801,"p99_latency_ms":3004,"time_to_first_token_ms":406,"total_runtime_ms":2546,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0252,"confidence_interval":0.9499,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2649,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2272,"p95_latency_ms":2777,"p99_latency_ms":2979,"time_to_first_token_ms":437,"total_runtime_ms":2525,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0279,"confidence_interval":0.9307,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7925,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2261,"p95_latency_ms":2764,"p99_latency_ms":2965,"time_to_first_token_ms":428,"total_runtime_ms":2512,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0284,"confidence_interval":0.9389,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7819,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2280,"p95_latency_ms":2786,"p99_latency_ms":2989,"time_to_first_token_ms":409,"total_runtime_ms":2533,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0256,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.164,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2322,"p95_latency_ms":2838,"p99_latency_ms":3044,"time_to_first_token_ms":422,"total_runtime_ms":2580,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0174,"confidence_interval":0.9369,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5627,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2318,"p95_latency_ms":2834,"p99_latency_ms":3040,"time_to_first_token_ms":437,"total_runtime_ms":2576,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0113,"confidence_interval":0.9321,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8077,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2252,"p95_latency_ms":2753,"p99_latency_ms":2953,"time_to_first_token_ms":430,"total_runtime_ms":2502,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0179,"confidence_interval":0.9325,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8381,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1274,"p95_latency_ms":1557,"p99_latency_ms":1671,"time_to_first_token_ms":385,"total_runtime_ms":1416,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0132,"confidence_interval":0.9413,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0777,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1221,"p95_latency_ms":1493,"p99_latency_ms":1601,"time_to_first_token_ms":377,"total_runtime_ms":1357,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0181,"confidence_interval":0.9351,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8359,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1290,"p95_latency_ms":1577,"p99_latency_ms":1692,"time_to_first_token_ms":395,"total_runtime_ms":1434,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0174,"confidence_interval":0.9429,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3238,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1272,"p95_latency_ms":1554,"p99_latency_ms":1667,"time_to_first_token_ms":394,"total_runtime_ms":1413,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0163,"confidence_interval":0.949,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5454,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1283,"p95_latency_ms":1568,"p99_latency_ms":1682,"time_to_first_token_ms":375,"total_runtime_ms":1425,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0245,"confidence_interval":0.9382,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3597,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1220,"p95_latency_ms":1491,"p99_latency_ms":1600,"time_to_first_token_ms":379,"total_runtime_ms":1356,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0278,"confidence_interval":0.9328,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4591,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1273,"p95_latency_ms":1556,"p99_latency_ms":1670,"time_to_first_token_ms":387,"total_runtime_ms":1415,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0265,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7211,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1218,"p95_latency_ms":1488,"p99_latency_ms":1597,"time_to_first_token_ms":362,"total_runtime_ms":1353,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0104,"confidence_interval":0.9485,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3381,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1223,"p95_latency_ms":1495,"p99_latency_ms":1604,"time_to_first_token_ms":395,"total_runtime_ms":1359,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.023,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2893,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1288,"p95_latency_ms":1574,"p99_latency_ms":1688,"time_to_first_token_ms":393,"total_runtime_ms":1431,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0247,"confidence_interval":0.9435,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6103,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3334,"p95_latency_ms":4075,"p99_latency_ms":4371,"time_to_first_token_ms":581,"total_runtime_ms":3704,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0112,"confidence_interval":0.9444,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8305,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3342,"p95_latency_ms":4084,"p99_latency_ms":4381,"time_to_first_token_ms":592,"total_runtime_ms":3713,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0247,"confidence_interval":0.9401,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6669,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3316,"p95_latency_ms":4052,"p99_latency_ms":4347,"time_to_first_token_ms":586,"total_runtime_ms":3684,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0288,"confidence_interval":0.9479,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1924,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3366,"p95_latency_ms":4113,"p99_latency_ms":4413,"time_to_first_token_ms":591,"total_runtime_ms":3740,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0177,"confidence_interval":0.9386,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6659,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3321,"p95_latency_ms":4058,"p99_latency_ms":4354,"time_to_first_token_ms":589,"total_runtime_ms":3690,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0254,"confidence_interval":0.9455,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4745,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3401,"p95_latency_ms":4157,"p99_latency_ms":4459,"time_to_first_token_ms":596,"total_runtime_ms":3779,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0259,"confidence_interval":0.9458,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0549,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3364,"p95_latency_ms":4112,"p99_latency_ms":4411,"time_to_first_token_ms":603,"total_runtime_ms":3738,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0123,"confidence_interval":0.9441,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2569,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3320,"p95_latency_ms":4058,"p99_latency_ms":4353,"time_to_first_token_ms":589,"total_runtime_ms":3689,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0153,"confidence_interval":0.9468,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9371,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3365,"p95_latency_ms":4113,"p99_latency_ms":4412,"time_to_first_token_ms":572,"total_runtime_ms":3739,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0229,"confidence_interval":0.9323,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9068,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3310,"p95_latency_ms":4046,"p99_latency_ms":4340,"time_to_first_token_ms":574,"total_runtime_ms":3678,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0189,"confidence_interval":0.9449,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6905,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1809,"p95_latency_ms":2211,"p99_latency_ms":2372,"time_to_first_token_ms":544,"total_runtime_ms":2010,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0147,"confidence_interval":0.9301,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6733,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1883,"p95_latency_ms":2302,"p99_latency_ms":2469,"time_to_first_token_ms":535,"total_runtime_ms":2092,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0185,"confidence_interval":0.9462,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4136,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1818,"p95_latency_ms":2222,"p99_latency_ms":2383,"time_to_first_token_ms":559,"total_runtime_ms":2020,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.023,"confidence_interval":0.9427,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.306,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1839,"p95_latency_ms":2248,"p99_latency_ms":2412,"time_to_first_token_ms":538,"total_runtime_ms":2044,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9315,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0175,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1839,"p95_latency_ms":2248,"p99_latency_ms":2412,"time_to_first_token_ms":530,"total_runtime_ms":2044,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0108,"confidence_interval":0.9464,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4399,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1832,"p95_latency_ms":2239,"p99_latency_ms":2402,"time_to_first_token_ms":543,"total_runtime_ms":2036,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.015,"confidence_interval":0.9459,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9972,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1801,"p95_latency_ms":2202,"p99_latency_ms":2362,"time_to_first_token_ms":547,"total_runtime_ms":2001,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0297,"confidence_interval":0.9491,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.345,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1875,"p95_latency_ms":2292,"p99_latency_ms":2459,"time_to_first_token_ms":538,"total_runtime_ms":2084,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0179,"confidence_interval":0.9364,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7576,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1872,"p95_latency_ms":2288,"p99_latency_ms":2454,"time_to_first_token_ms":557,"total_runtime_ms":2080,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0197,"confidence_interval":0.9319,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4124,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|low|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1865,"p95_latency_ms":2280,"p99_latency_ms":2446,"time_to_first_token_ms":528,"total_runtime_ms":2073,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0171,"confidence_interval":0.9444,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5445,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":627,"p95_latency_ms":766,"p99_latency_ms":822,"time_to_first_token_ms":168,"total_runtime_ms":697,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0176,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2311,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":709,"p95_latency_ms":867,"p99_latency_ms":930,"time_to_first_token_ms":180,"total_runtime_ms":788,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1023,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":725,"p95_latency_ms":886,"p99_latency_ms":950,"time_to_first_token_ms":189,"total_runtime_ms":805,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0117,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.88,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":721,"p95_latency_ms":881,"p99_latency_ms":945,"time_to_first_token_ms":178,"total_runtime_ms":801,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7777,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":712,"p95_latency_ms":871,"p99_latency_ms":934,"time_to_first_token_ms":178,"total_runtime_ms":792,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0139,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1599,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":931,"time_to_first_token_ms":156,"total_runtime_ms":789,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0108,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.208,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":709,"p95_latency_ms":867,"p99_latency_ms":930,"time_to_first_token_ms":159,"total_runtime_ms":788,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0199,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4498,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":867,"time_to_first_token_ms":187,"total_runtime_ms":734,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0229,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8188,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":645,"p95_latency_ms":789,"p99_latency_ms":846,"time_to_first_token_ms":190,"total_runtime_ms":717,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0115,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6646,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":692,"p95_latency_ms":845,"p99_latency_ms":907,"time_to_first_token_ms":157,"total_runtime_ms":769,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0192,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9295,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":382,"p95_latency_ms":467,"p99_latency_ms":501,"time_to_first_token_ms":113,"total_runtime_ms":424,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0176,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7152,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":495,"time_to_first_token_ms":137,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0197,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9235,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":348,"p95_latency_ms":426,"p99_latency_ms":457,"time_to_first_token_ms":140,"total_runtime_ms":387,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0269,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0932,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":409,"p95_latency_ms":500,"p99_latency_ms":537,"time_to_first_token_ms":113,"total_runtime_ms":455,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0198,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6983,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":404,"p95_latency_ms":494,"p99_latency_ms":529,"time_to_first_token_ms":124,"total_runtime_ms":449,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0249,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.506,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":405,"p95_latency_ms":495,"p99_latency_ms":531,"time_to_first_token_ms":126,"total_runtime_ms":450,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0125,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1293,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":414,"p95_latency_ms":507,"p99_latency_ms":543,"time_to_first_token_ms":114,"total_runtime_ms":461,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0229,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5999,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":371,"p95_latency_ms":454,"p99_latency_ms":487,"time_to_first_token_ms":113,"total_runtime_ms":413,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0246,"confidence_interval":0.9468,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4341,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":483,"time_to_first_token_ms":137,"total_runtime_ms":409,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0286,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1871,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":429,"p95_latency_ms":524,"p99_latency_ms":562,"time_to_first_token_ms":132,"total_runtime_ms":476,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0141,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1986,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":911,"p95_latency_ms":1113,"p99_latency_ms":1194,"time_to_first_token_ms":191,"total_runtime_ms":1012,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0154,"confidence_interval":0.9427,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6276,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":912,"p95_latency_ms":1115,"p99_latency_ms":1196,"time_to_first_token_ms":196,"total_runtime_ms":1014,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0254,"confidence_interval":0.9457,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5626,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":896,"p95_latency_ms":1095,"p99_latency_ms":1175,"time_to_first_token_ms":203,"total_runtime_ms":996,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0118,"confidence_interval":0.9398,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0959,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":869,"p95_latency_ms":1062,"p99_latency_ms":1139,"time_to_first_token_ms":221,"total_runtime_ms":965,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0199,"confidence_interval":0.9399,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0618,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":874,"p95_latency_ms":1069,"p99_latency_ms":1146,"time_to_first_token_ms":223,"total_runtime_ms":971,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0287,"confidence_interval":0.9439,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0709,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":936,"p95_latency_ms":1144,"p99_latency_ms":1228,"time_to_first_token_ms":218,"total_runtime_ms":1040,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0265,"confidence_interval":0.9361,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1153,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":895,"p95_latency_ms":1094,"p99_latency_ms":1174,"time_to_first_token_ms":221,"total_runtime_ms":995,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0176,"confidence_interval":0.9384,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5547,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":861,"p95_latency_ms":1053,"p99_latency_ms":1129,"time_to_first_token_ms":192,"total_runtime_ms":957,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0268,"confidence_interval":0.9385,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7911,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":897,"p95_latency_ms":1097,"p99_latency_ms":1177,"time_to_first_token_ms":216,"total_runtime_ms":997,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0267,"confidence_interval":0.9309,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.058,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":889,"p95_latency_ms":1087,"p99_latency_ms":1166,"time_to_first_token_ms":209,"total_runtime_ms":988,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6944,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":509,"p95_latency_ms":622,"p99_latency_ms":667,"time_to_first_token_ms":178,"total_runtime_ms":566,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0165,"confidence_interval":0.9345,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7587,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":505,"p95_latency_ms":618,"p99_latency_ms":663,"time_to_first_token_ms":160,"total_runtime_ms":562,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.014,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2374,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":162,"total_runtime_ms":586,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0192,"confidence_interval":0.9459,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9303,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":471,"p95_latency_ms":575,"p99_latency_ms":617,"time_to_first_token_ms":147,"total_runtime_ms":523,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0242,"confidence_interval":0.9386,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1559,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":491,"p95_latency_ms":601,"p99_latency_ms":644,"time_to_first_token_ms":159,"total_runtime_ms":546,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0259,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8301,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":540,"p95_latency_ms":660,"p99_latency_ms":708,"time_to_first_token_ms":158,"total_runtime_ms":600,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0196,"confidence_interval":0.941,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.966,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":157,"total_runtime_ms":536,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0127,"confidence_interval":0.9355,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0005,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":503,"p95_latency_ms":614,"p99_latency_ms":659,"time_to_first_token_ms":175,"total_runtime_ms":558,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0135,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.162,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":492,"p95_latency_ms":601,"p99_latency_ms":645,"time_to_first_token_ms":146,"total_runtime_ms":547,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.012,"confidence_interval":0.9416,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6286,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":544,"p95_latency_ms":665,"p99_latency_ms":714,"time_to_first_token_ms":146,"total_runtime_ms":605,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0181,"confidence_interval":0.9441,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7361,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1755,"p95_latency_ms":2144,"p99_latency_ms":2300,"time_to_first_token_ms":323,"total_runtime_ms":1950,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0225,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9006,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1696,"p95_latency_ms":2073,"p99_latency_ms":2224,"time_to_first_token_ms":320,"total_runtime_ms":1885,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0171,"confidence_interval":0.9492,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2733,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1703,"p95_latency_ms":2082,"p99_latency_ms":2233,"time_to_first_token_ms":334,"total_runtime_ms":1893,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0196,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8425,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1757,"p95_latency_ms":2147,"p99_latency_ms":2304,"time_to_first_token_ms":336,"total_runtime_ms":1952,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0115,"confidence_interval":0.9331,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4597,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1726,"p95_latency_ms":2109,"p99_latency_ms":2263,"time_to_first_token_ms":323,"total_runtime_ms":1918,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0293,"confidence_interval":0.9355,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4428,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1772,"p95_latency_ms":2166,"p99_latency_ms":2324,"time_to_first_token_ms":340,"total_runtime_ms":1969,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0159,"confidence_interval":0.9498,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1963,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1774,"p95_latency_ms":2168,"p99_latency_ms":2325,"time_to_first_token_ms":321,"total_runtime_ms":1971,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.028,"confidence_interval":0.946,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7472,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1722,"p95_latency_ms":2104,"p99_latency_ms":2257,"time_to_first_token_ms":340,"total_runtime_ms":1913,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0155,"confidence_interval":0.9433,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3089,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1763,"p95_latency_ms":2154,"p99_latency_ms":2311,"time_to_first_token_ms":338,"total_runtime_ms":1958,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0144,"confidence_interval":0.933,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9741,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1728,"p95_latency_ms":2112,"p99_latency_ms":2266,"time_to_first_token_ms":321,"total_runtime_ms":1920,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0116,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1395,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":993,"p95_latency_ms":1213,"p99_latency_ms":1302,"time_to_first_token_ms":290,"total_runtime_ms":1103,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0158,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8843,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":966,"p95_latency_ms":1181,"p99_latency_ms":1267,"time_to_first_token_ms":291,"total_runtime_ms":1074,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0292,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.216,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":930,"p95_latency_ms":1136,"p99_latency_ms":1219,"time_to_first_token_ms":282,"total_runtime_ms":1033,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0288,"confidence_interval":0.9448,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6284,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":931,"p95_latency_ms":1138,"p99_latency_ms":1221,"time_to_first_token_ms":281,"total_runtime_ms":1035,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0265,"confidence_interval":0.9339,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.549,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":937,"p95_latency_ms":1146,"p99_latency_ms":1229,"time_to_first_token_ms":305,"total_runtime_ms":1042,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0262,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1803,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":943,"p95_latency_ms":1153,"p99_latency_ms":1237,"time_to_first_token_ms":275,"total_runtime_ms":1048,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0215,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6845,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":942,"p95_latency_ms":1151,"p99_latency_ms":1235,"time_to_first_token_ms":289,"total_runtime_ms":1047,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0129,"confidence_interval":0.9433,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.361,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":950,"p95_latency_ms":1162,"p99_latency_ms":1246,"time_to_first_token_ms":302,"total_runtime_ms":1056,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0225,"confidence_interval":0.9382,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9476,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":997,"p95_latency_ms":1219,"p99_latency_ms":1308,"time_to_first_token_ms":274,"total_runtime_ms":1108,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0191,"confidence_interval":0.9477,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9674,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":959,"p95_latency_ms":1172,"p99_latency_ms":1258,"time_to_first_token_ms":274,"total_runtime_ms":1066,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0268,"confidence_interval":0.9332,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9802,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2503,"p95_latency_ms":3059,"p99_latency_ms":3282,"time_to_first_token_ms":456,"total_runtime_ms":2781,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0179,"confidence_interval":0.9436,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7874,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2575,"p95_latency_ms":3148,"p99_latency_ms":3377,"time_to_first_token_ms":456,"total_runtime_ms":2861,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0135,"confidence_interval":0.9379,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3498,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2502,"p95_latency_ms":3058,"p99_latency_ms":3280,"time_to_first_token_ms":467,"total_runtime_ms":2780,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0115,"confidence_interval":0.936,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3074,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2480,"p95_latency_ms":3031,"p99_latency_ms":3252,"time_to_first_token_ms":464,"total_runtime_ms":2756,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0276,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9644,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2549,"p95_latency_ms":3115,"p99_latency_ms":3342,"time_to_first_token_ms":449,"total_runtime_ms":2832,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0151,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0748,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2575,"p95_latency_ms":3147,"p99_latency_ms":3376,"time_to_first_token_ms":471,"total_runtime_ms":2861,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0164,"confidence_interval":0.9302,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4719,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2475,"p95_latency_ms":3026,"p99_latency_ms":3246,"time_to_first_token_ms":445,"total_runtime_ms":2751,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0144,"confidence_interval":0.9371,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7263,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2500,"p95_latency_ms":3055,"p99_latency_ms":3278,"time_to_first_token_ms":465,"total_runtime_ms":2778,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0282,"confidence_interval":0.9404,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5346,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2548,"p95_latency_ms":3114,"p99_latency_ms":3340,"time_to_first_token_ms":470,"total_runtime_ms":2831,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.016,"confidence_interval":0.9333,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8858,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2548,"p95_latency_ms":3115,"p99_latency_ms":3341,"time_to_first_token_ms":460,"total_runtime_ms":2832,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0204,"confidence_interval":0.9311,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9968,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1419,"p95_latency_ms":1734,"p99_latency_ms":1860,"time_to_first_token_ms":410,"total_runtime_ms":1576,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0263,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1828,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1373,"p95_latency_ms":1678,"p99_latency_ms":1800,"time_to_first_token_ms":425,"total_runtime_ms":1525,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0119,"confidence_interval":0.9314,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2586,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1373,"p95_latency_ms":1678,"p99_latency_ms":1800,"time_to_first_token_ms":423,"total_runtime_ms":1525,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0167,"confidence_interval":0.9315,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4368,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1352,"p95_latency_ms":1652,"p99_latency_ms":1773,"time_to_first_token_ms":405,"total_runtime_ms":1502,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0227,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9397,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1425,"p95_latency_ms":1742,"p99_latency_ms":1868,"time_to_first_token_ms":428,"total_runtime_ms":1583,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9342,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5942,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1435,"p95_latency_ms":1754,"p99_latency_ms":1882,"time_to_first_token_ms":422,"total_runtime_ms":1595,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0174,"confidence_interval":0.9393,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1517,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1424,"p95_latency_ms":1740,"p99_latency_ms":1866,"time_to_first_token_ms":431,"total_runtime_ms":1582,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0251,"confidence_interval":0.9379,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2952,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1402,"p95_latency_ms":1713,"p99_latency_ms":1838,"time_to_first_token_ms":424,"total_runtime_ms":1557,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0175,"confidence_interval":0.932,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4199,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1420,"p95_latency_ms":1735,"p99_latency_ms":1861,"time_to_first_token_ms":403,"total_runtime_ms":1577,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0262,"confidence_interval":0.9466,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2273,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1410,"p95_latency_ms":1724,"p99_latency_ms":1849,"time_to_first_token_ms":410,"total_runtime_ms":1567,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0185,"confidence_interval":0.9404,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3904,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":495,"p95_latency_ms":606,"p99_latency_ms":650,"time_to_first_token_ms":162,"total_runtime_ms":551,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0157,"confidence_interval":0.9351,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.128,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":487,"p95_latency_ms":595,"p99_latency_ms":638,"time_to_first_token_ms":166,"total_runtime_ms":541,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3746,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":531,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":155,"total_runtime_ms":591,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0274,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5393,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":494,"p95_latency_ms":603,"p99_latency_ms":647,"time_to_first_token_ms":139,"total_runtime_ms":549,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0237,"confidence_interval":0.9402,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8236,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":158,"total_runtime_ms":585,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.021,"confidence_interval":0.9424,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5541,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":564,"p95_latency_ms":690,"p99_latency_ms":740,"time_to_first_token_ms":141,"total_runtime_ms":627,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0104,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9038,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":490,"p95_latency_ms":599,"p99_latency_ms":642,"time_to_first_token_ms":140,"total_runtime_ms":544,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0299,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3965,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":525,"p95_latency_ms":642,"p99_latency_ms":688,"time_to_first_token_ms":166,"total_runtime_ms":583,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7981,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":551,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":146,"total_runtime_ms":613,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2916,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":533,"p95_latency_ms":651,"p99_latency_ms":699,"time_to_first_token_ms":166,"total_runtime_ms":592,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0266,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5162,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":306,"p95_latency_ms":374,"p99_latency_ms":402,"time_to_first_token_ms":120,"total_runtime_ms":340,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0298,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4308,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":320,"p95_latency_ms":391,"p99_latency_ms":420,"time_to_first_token_ms":102,"total_runtime_ms":356,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0159,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5567,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":271,"p95_latency_ms":331,"p99_latency_ms":355,"time_to_first_token_ms":120,"total_runtime_ms":301,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0244,"confidence_interval":0.9397,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.353,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":334,"p95_latency_ms":409,"p99_latency_ms":438,"time_to_first_token_ms":119,"total_runtime_ms":371,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0155,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0305,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":329,"p95_latency_ms":403,"p99_latency_ms":432,"time_to_first_token_ms":104,"total_runtime_ms":366,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0165,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0124,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":327,"p95_latency_ms":399,"p99_latency_ms":428,"time_to_first_token_ms":103,"total_runtime_ms":363,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3318,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":353,"p95_latency_ms":432,"p99_latency_ms":463,"time_to_first_token_ms":100,"total_runtime_ms":393,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0293,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.07,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":253,"p95_latency_ms":309,"p99_latency_ms":331,"time_to_first_token_ms":94,"total_runtime_ms":281,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5385,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":360,"p95_latency_ms":440,"p99_latency_ms":472,"time_to_first_token_ms":89,"total_runtime_ms":400,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0105,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.68,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":312,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":94,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0185,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8999,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":635,"p95_latency_ms":777,"p99_latency_ms":833,"time_to_first_token_ms":178,"total_runtime_ms":706,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0178,"confidence_interval":0.9306,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2323,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":666,"p95_latency_ms":813,"p99_latency_ms":873,"time_to_first_token_ms":179,"total_runtime_ms":739,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0299,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5725,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":711,"p95_latency_ms":869,"p99_latency_ms":932,"time_to_first_token_ms":157,"total_runtime_ms":790,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0283,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2162,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":634,"p95_latency_ms":774,"p99_latency_ms":831,"time_to_first_token_ms":175,"total_runtime_ms":704,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0248,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9566,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":948,"time_to_first_token_ms":178,"total_runtime_ms":804,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0182,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0563,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":637,"p95_latency_ms":778,"p99_latency_ms":835,"time_to_first_token_ms":181,"total_runtime_ms":708,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0167,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7639,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":706,"p95_latency_ms":863,"p99_latency_ms":926,"time_to_first_token_ms":167,"total_runtime_ms":785,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0281,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.413,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":720,"p95_latency_ms":880,"p99_latency_ms":944,"time_to_first_token_ms":184,"total_runtime_ms":800,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0291,"confidence_interval":0.9478,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1029,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":708,"p95_latency_ms":866,"p99_latency_ms":929,"time_to_first_token_ms":165,"total_runtime_ms":787,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0225,"confidence_interval":0.931,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4937,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":696,"p95_latency_ms":851,"p99_latency_ms":912,"time_to_first_token_ms":180,"total_runtime_ms":773,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0123,"confidence_interval":0.944,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0742,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":352,"p95_latency_ms":431,"p99_latency_ms":462,"time_to_first_token_ms":147,"total_runtime_ms":392,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0159,"confidence_interval":0.932,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7372,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":414,"p95_latency_ms":506,"p99_latency_ms":542,"time_to_first_token_ms":142,"total_runtime_ms":460,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0293,"confidence_interval":0.9466,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5605,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":413,"p95_latency_ms":505,"p99_latency_ms":541,"time_to_first_token_ms":140,"total_runtime_ms":459,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.024,"confidence_interval":0.942,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2595,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":360,"p95_latency_ms":441,"p99_latency_ms":473,"time_to_first_token_ms":114,"total_runtime_ms":401,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0154,"confidence_interval":0.9413,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6507,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":397,"p95_latency_ms":485,"p99_latency_ms":520,"time_to_first_token_ms":121,"total_runtime_ms":441,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0192,"confidence_interval":0.9331,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1233,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":381,"p95_latency_ms":466,"p99_latency_ms":499,"time_to_first_token_ms":126,"total_runtime_ms":423,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0217,"confidence_interval":0.9406,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3652,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":372,"p95_latency_ms":454,"p99_latency_ms":487,"time_to_first_token_ms":142,"total_runtime_ms":413,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0259,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9801,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":398,"p95_latency_ms":486,"p99_latency_ms":522,"time_to_first_token_ms":125,"total_runtime_ms":442,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0299,"confidence_interval":0.9477,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8285,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":350,"p95_latency_ms":428,"p99_latency_ms":460,"time_to_first_token_ms":134,"total_runtime_ms":389,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0141,"confidence_interval":0.9468,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9425,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":127,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.9342,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0833,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1299,"p95_latency_ms":1588,"p99_latency_ms":1703,"time_to_first_token_ms":278,"total_runtime_ms":1443,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.011,"confidence_interval":0.943,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6855,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1325,"p95_latency_ms":1619,"p99_latency_ms":1737,"time_to_first_token_ms":279,"total_runtime_ms":1472,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0116,"confidence_interval":0.9491,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4965,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1344,"p95_latency_ms":1642,"p99_latency_ms":1762,"time_to_first_token_ms":283,"total_runtime_ms":1493,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0138,"confidence_interval":0.9344,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1196,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1312,"p95_latency_ms":1603,"p99_latency_ms":1720,"time_to_first_token_ms":267,"total_runtime_ms":1457,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.029,"confidence_interval":0.939,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0542,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1270,"p95_latency_ms":1553,"p99_latency_ms":1666,"time_to_first_token_ms":275,"total_runtime_ms":1412,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.014,"confidence_interval":0.9348,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8187,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1341,"p95_latency_ms":1639,"p99_latency_ms":1758,"time_to_first_token_ms":272,"total_runtime_ms":1490,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0278,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1595,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1253,"p95_latency_ms":1531,"p99_latency_ms":1643,"time_to_first_token_ms":276,"total_runtime_ms":1392,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0192,"confidence_interval":0.9301,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2754,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1307,"p95_latency_ms":1598,"p99_latency_ms":1714,"time_to_first_token_ms":262,"total_runtime_ms":1453,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0256,"confidence_interval":0.9326,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8707,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1288,"p95_latency_ms":1574,"p99_latency_ms":1688,"time_to_first_token_ms":266,"total_runtime_ms":1431,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0125,"confidence_interval":0.938,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4853,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1325,"p95_latency_ms":1619,"p99_latency_ms":1737,"time_to_first_token_ms":279,"total_runtime_ms":1472,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.017,"confidence_interval":0.9376,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.071,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":727,"p95_latency_ms":889,"p99_latency_ms":953,"time_to_first_token_ms":236,"total_runtime_ms":808,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0217,"confidence_interval":0.9454,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.389,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":753,"p95_latency_ms":920,"p99_latency_ms":987,"time_to_first_token_ms":239,"total_runtime_ms":836,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.9497,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8982,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":787,"p95_latency_ms":962,"p99_latency_ms":1032,"time_to_first_token_ms":227,"total_runtime_ms":875,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0261,"confidence_interval":0.9459,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5323,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":733,"p95_latency_ms":896,"p99_latency_ms":961,"time_to_first_token_ms":214,"total_runtime_ms":814,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0181,"confidence_interval":0.943,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0393,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":714,"p95_latency_ms":872,"p99_latency_ms":936,"time_to_first_token_ms":239,"total_runtime_ms":793,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0278,"confidence_interval":0.9471,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3519,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":693,"p95_latency_ms":847,"p99_latency_ms":908,"time_to_first_token_ms":217,"total_runtime_ms":770,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0236,"confidence_interval":0.9371,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3749,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":759,"p95_latency_ms":928,"p99_latency_ms":996,"time_to_first_token_ms":229,"total_runtime_ms":844,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0195,"confidence_interval":0.9327,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4015,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":705,"p95_latency_ms":861,"p99_latency_ms":924,"time_to_first_token_ms":224,"total_runtime_ms":783,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0247,"confidence_interval":0.9377,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9414,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":751,"p95_latency_ms":918,"p99_latency_ms":985,"time_to_first_token_ms":233,"total_runtime_ms":835,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0193,"confidence_interval":0.9404,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4302,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1003,"time_to_first_token_ms":218,"total_runtime_ms":850,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.9494,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9373,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1948,"p95_latency_ms":2381,"p99_latency_ms":2554,"time_to_first_token_ms":382,"total_runtime_ms":2164,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0157,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.185,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1876,"p95_latency_ms":2293,"p99_latency_ms":2460,"time_to_first_token_ms":364,"total_runtime_ms":2085,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.024,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3469,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1961,"p95_latency_ms":2397,"p99_latency_ms":2571,"time_to_first_token_ms":362,"total_runtime_ms":2179,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0167,"confidence_interval":0.9375,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2054,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1904,"p95_latency_ms":2328,"p99_latency_ms":2497,"time_to_first_token_ms":355,"total_runtime_ms":2116,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0219,"confidence_interval":0.9429,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1475,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1943,"p95_latency_ms":2374,"p99_latency_ms":2547,"time_to_first_token_ms":348,"total_runtime_ms":2158,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0189,"confidence_interval":0.9472,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7095,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1887,"p95_latency_ms":2307,"p99_latency_ms":2475,"time_to_first_token_ms":373,"total_runtime_ms":2097,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0201,"confidence_interval":0.9415,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6397,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1934,"p95_latency_ms":2364,"p99_latency_ms":2536,"time_to_first_token_ms":352,"total_runtime_ms":2149,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0229,"confidence_interval":0.9336,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.047,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1871,"p95_latency_ms":2286,"p99_latency_ms":2453,"time_to_first_token_ms":365,"total_runtime_ms":2079,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0165,"confidence_interval":0.9456,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8059,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1858,"p95_latency_ms":2271,"p99_latency_ms":2436,"time_to_first_token_ms":371,"total_runtime_ms":2064,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0201,"confidence_interval":0.9461,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0614,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1915,"p95_latency_ms":2341,"p99_latency_ms":2511,"time_to_first_token_ms":353,"total_runtime_ms":2128,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0208,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4169,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1117,"p95_latency_ms":1365,"p99_latency_ms":1464,"time_to_first_token_ms":306,"total_runtime_ms":1241,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0259,"confidence_interval":0.9471,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6431,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1344,"time_to_first_token_ms":312,"total_runtime_ms":1139,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0169,"confidence_interval":0.9328,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8527,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1015,"p95_latency_ms":1241,"p99_latency_ms":1331,"time_to_first_token_ms":317,"total_runtime_ms":1128,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0115,"confidence_interval":0.936,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9903,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1071,"p95_latency_ms":1309,"p99_latency_ms":1404,"time_to_first_token_ms":324,"total_runtime_ms":1190,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0161,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.25,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1046,"p95_latency_ms":1278,"p99_latency_ms":1371,"time_to_first_token_ms":322,"total_runtime_ms":1162,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0189,"confidence_interval":0.9321,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1456,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1034,"p95_latency_ms":1264,"p99_latency_ms":1356,"time_to_first_token_ms":332,"total_runtime_ms":1149,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.017,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0417,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1087,"p95_latency_ms":1329,"p99_latency_ms":1426,"time_to_first_token_ms":323,"total_runtime_ms":1208,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0264,"confidence_interval":0.933,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.758,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1096,"p95_latency_ms":1340,"p99_latency_ms":1437,"time_to_first_token_ms":322,"total_runtime_ms":1218,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0178,"confidence_interval":0.9433,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5174,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1041,"p95_latency_ms":1273,"p99_latency_ms":1365,"time_to_first_token_ms":331,"total_runtime_ms":1157,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0129,"confidence_interval":0.9499,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7799,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1038,"p95_latency_ms":1268,"p99_latency_ms":1361,"time_to_first_token_ms":311,"total_runtime_ms":1153,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0151,"confidence_interval":0.9425,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8522,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":453,"p95_latency_ms":554,"p99_latency_ms":594,"time_to_first_token_ms":128,"total_runtime_ms":503,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0267,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4055,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":577,"time_to_first_token_ms":140,"total_runtime_ms":489,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0114,"confidence_interval":0.9397,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5957,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":468,"p95_latency_ms":572,"p99_latency_ms":613,"time_to_first_token_ms":153,"total_runtime_ms":520,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0163,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5877,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":402,"p95_latency_ms":491,"p99_latency_ms":527,"time_to_first_token_ms":126,"total_runtime_ms":446,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0254,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.686,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":485,"p95_latency_ms":593,"p99_latency_ms":636,"time_to_first_token_ms":129,"total_runtime_ms":539,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0226,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7747,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":394,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":140,"total_runtime_ms":437,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0296,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3034,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":396,"p95_latency_ms":484,"p99_latency_ms":519,"time_to_first_token_ms":143,"total_runtime_ms":440,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0292,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5806,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":399,"p95_latency_ms":488,"p99_latency_ms":523,"time_to_first_token_ms":130,"total_runtime_ms":443,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0174,"confidence_interval":0.9414,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1285,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":438,"p95_latency_ms":535,"p99_latency_ms":574,"time_to_first_token_ms":123,"total_runtime_ms":486,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0146,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3897,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":150,"total_runtime_ms":456,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.03,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8565,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":258,"p95_latency_ms":315,"p99_latency_ms":338,"time_to_first_token_ms":107,"total_runtime_ms":286,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0231,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6543,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":277,"p95_latency_ms":339,"p99_latency_ms":364,"time_to_first_token_ms":109,"total_runtime_ms":308,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0139,"confidence_interval":0.9445,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6583,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":291,"p99_latency_ms":313,"time_to_first_token_ms":76,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0275,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.864,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":227,"p95_latency_ms":278,"p99_latency_ms":298,"time_to_first_token_ms":86,"total_runtime_ms":253,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0182,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3122,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":242,"p95_latency_ms":296,"p99_latency_ms":317,"time_to_first_token_ms":95,"total_runtime_ms":269,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0199,"confidence_interval":0.9348,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.877,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":227,"p95_latency_ms":277,"p99_latency_ms":297,"time_to_first_token_ms":109,"total_runtime_ms":252,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0259,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3726,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":291,"p99_latency_ms":312,"time_to_first_token_ms":81,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0298,"confidence_interval":0.9376,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1928,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":266,"p95_latency_ms":326,"p99_latency_ms":349,"time_to_first_token_ms":75,"total_runtime_ms":296,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.021,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2997,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":275,"p95_latency_ms":336,"p99_latency_ms":361,"time_to_first_token_ms":103,"total_runtime_ms":306,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0112,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2983,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":249,"p95_latency_ms":304,"p99_latency_ms":326,"time_to_first_token_ms":86,"total_runtime_ms":276,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0129,"confidence_interval":0.9331,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9272,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":552,"p95_latency_ms":675,"p99_latency_ms":724,"time_to_first_token_ms":146,"total_runtime_ms":614,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0164,"confidence_interval":0.9454,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4903,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":175,"total_runtime_ms":591,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0242,"confidence_interval":0.9429,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8396,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":141,"total_runtime_ms":585,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.018,"confidence_interval":0.9496,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6646,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":574,"p95_latency_ms":701,"p99_latency_ms":752,"time_to_first_token_ms":161,"total_runtime_ms":637,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0135,"confidence_interval":0.9372,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.262,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":580,"p95_latency_ms":709,"p99_latency_ms":760,"time_to_first_token_ms":168,"total_runtime_ms":644,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0186,"confidence_interval":0.9417,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3553,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":785,"time_to_first_token_ms":176,"total_runtime_ms":665,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0282,"confidence_interval":0.942,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2013,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":587,"p95_latency_ms":717,"p99_latency_ms":769,"time_to_first_token_ms":153,"total_runtime_ms":652,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.013,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8277,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":568,"p95_latency_ms":695,"p99_latency_ms":745,"time_to_first_token_ms":146,"total_runtime_ms":632,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0282,"confidence_interval":0.9379,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8683,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":626,"p95_latency_ms":765,"p99_latency_ms":821,"time_to_first_token_ms":173,"total_runtime_ms":695,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0228,"confidence_interval":0.9456,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0846,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":150,"total_runtime_ms":664,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.02,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5419,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":291,"p95_latency_ms":355,"p99_latency_ms":381,"time_to_first_token_ms":110,"total_runtime_ms":323,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0198,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1056,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":294,"p95_latency_ms":359,"p99_latency_ms":385,"time_to_first_token_ms":127,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0131,"confidence_interval":0.9398,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6943,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":357,"p95_latency_ms":437,"p99_latency_ms":468,"time_to_first_token_ms":113,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0155,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7722,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":377,"p99_latency_ms":404,"time_to_first_token_ms":120,"total_runtime_ms":343,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0179,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2904,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":406,"time_to_first_token_ms":97,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0154,"confidence_interval":0.9368,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9006,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":476,"p99_latency_ms":511,"time_to_first_token_ms":102,"total_runtime_ms":433,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0101,"confidence_interval":0.9318,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7471,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":321,"p95_latency_ms":392,"p99_latency_ms":421,"time_to_first_token_ms":104,"total_runtime_ms":356,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.025,"confidence_interval":0.9301,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1783,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":337,"p95_latency_ms":412,"p99_latency_ms":442,"time_to_first_token_ms":116,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0234,"confidence_interval":0.9326,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8928,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":347,"p95_latency_ms":425,"p99_latency_ms":456,"time_to_first_token_ms":98,"total_runtime_ms":386,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0298,"confidence_interval":0.9472,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9635,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":128,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0255,"confidence_interval":0.9495,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2341,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1095,"p95_latency_ms":1338,"p99_latency_ms":1436,"time_to_first_token_ms":233,"total_runtime_ms":1217,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0108,"confidence_interval":0.9437,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7021,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1090,"p95_latency_ms":1332,"p99_latency_ms":1429,"time_to_first_token_ms":223,"total_runtime_ms":1211,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.9462,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.301,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1060,"p95_latency_ms":1295,"p99_latency_ms":1390,"time_to_first_token_ms":251,"total_runtime_ms":1178,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0212,"confidence_interval":0.9316,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6798,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1049,"p95_latency_ms":1282,"p99_latency_ms":1375,"time_to_first_token_ms":243,"total_runtime_ms":1165,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0293,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7862,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1072,"p95_latency_ms":1310,"p99_latency_ms":1406,"time_to_first_token_ms":257,"total_runtime_ms":1191,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0281,"confidence_interval":0.935,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1477,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1067,"p95_latency_ms":1304,"p99_latency_ms":1398,"time_to_first_token_ms":235,"total_runtime_ms":1185,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0266,"confidence_interval":0.947,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0778,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1065,"p95_latency_ms":1302,"p99_latency_ms":1397,"time_to_first_token_ms":227,"total_runtime_ms":1184,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0262,"confidence_interval":0.9491,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7887,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1149,"p95_latency_ms":1404,"p99_latency_ms":1506,"time_to_first_token_ms":226,"total_runtime_ms":1277,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0127,"confidence_interval":0.9437,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7938,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1076,"p95_latency_ms":1316,"p99_latency_ms":1411,"time_to_first_token_ms":232,"total_runtime_ms":1196,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0241,"confidence_interval":0.9306,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5017,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1051,"p95_latency_ms":1284,"p99_latency_ms":1377,"time_to_first_token_ms":242,"total_runtime_ms":1167,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0222,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7251,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":656,"p95_latency_ms":802,"p99_latency_ms":860,"time_to_first_token_ms":188,"total_runtime_ms":729,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0126,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6403,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":592,"p95_latency_ms":724,"p99_latency_ms":777,"time_to_first_token_ms":190,"total_runtime_ms":658,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0222,"confidence_interval":0.947,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7498,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":645,"p95_latency_ms":788,"p99_latency_ms":845,"time_to_first_token_ms":210,"total_runtime_ms":716,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.019,"confidence_interval":0.9444,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2844,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":630,"p95_latency_ms":770,"p99_latency_ms":826,"time_to_first_token_ms":182,"total_runtime_ms":700,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0122,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2327,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":665,"p95_latency_ms":813,"p99_latency_ms":872,"time_to_first_token_ms":210,"total_runtime_ms":739,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0158,"confidence_interval":0.9437,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4242,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":651,"p95_latency_ms":796,"p99_latency_ms":854,"time_to_first_token_ms":189,"total_runtime_ms":724,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0204,"confidence_interval":0.9444,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9146,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":201,"total_runtime_ms":663,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0119,"confidence_interval":0.9489,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5599,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":669,"p95_latency_ms":818,"p99_latency_ms":878,"time_to_first_token_ms":189,"total_runtime_ms":744,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9349,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5579,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":803,"time_to_first_token_ms":182,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0294,"confidence_interval":0.946,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.103,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:7b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":611,"p95_latency_ms":747,"p99_latency_ms":801,"time_to_first_token_ms":197,"total_runtime_ms":679,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0112,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9802,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1562,"p95_latency_ms":1909,"p99_latency_ms":2047,"time_to_first_token_ms":325,"total_runtime_ms":1735,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0191,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0564,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1607,"p95_latency_ms":1964,"p99_latency_ms":2107,"time_to_first_token_ms":330,"total_runtime_ms":1786,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0133,"confidence_interval":0.9307,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3696,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1631,"p95_latency_ms":1994,"p99_latency_ms":2139,"time_to_first_token_ms":316,"total_runtime_ms":1813,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0118,"confidence_interval":0.9466,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4691,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1559,"p95_latency_ms":1905,"p99_latency_ms":2043,"time_to_first_token_ms":333,"total_runtime_ms":1732,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0245,"confidence_interval":0.9358,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.374,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1545,"p95_latency_ms":1889,"p99_latency_ms":2026,"time_to_first_token_ms":303,"total_runtime_ms":1717,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0226,"confidence_interval":0.9384,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4744,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1578,"p95_latency_ms":1929,"p99_latency_ms":2069,"time_to_first_token_ms":334,"total_runtime_ms":1753,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0175,"confidence_interval":0.9422,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7826,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1606,"p95_latency_ms":1963,"p99_latency_ms":2106,"time_to_first_token_ms":301,"total_runtime_ms":1784,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0102,"confidence_interval":0.9486,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1441,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1633,"p95_latency_ms":1995,"p99_latency_ms":2141,"time_to_first_token_ms":325,"total_runtime_ms":1814,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.944,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8866,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1628,"p95_latency_ms":1990,"p99_latency_ms":2135,"time_to_first_token_ms":301,"total_runtime_ms":1809,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0176,"confidence_interval":0.9465,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7631,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1558,"p95_latency_ms":1904,"p99_latency_ms":2042,"time_to_first_token_ms":323,"total_runtime_ms":1731,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0204,"confidence_interval":0.9369,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6478,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":861,"p95_latency_ms":1053,"p99_latency_ms":1129,"time_to_first_token_ms":289,"total_runtime_ms":957,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0248,"confidence_interval":0.9431,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1148,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":905,"p95_latency_ms":1107,"p99_latency_ms":1187,"time_to_first_token_ms":279,"total_runtime_ms":1006,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0113,"confidence_interval":0.9465,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2407,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":884,"p95_latency_ms":1080,"p99_latency_ms":1159,"time_to_first_token_ms":283,"total_runtime_ms":982,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0212,"confidence_interval":0.9419,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7881,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":931,"p95_latency_ms":1138,"p99_latency_ms":1221,"time_to_first_token_ms":278,"total_runtime_ms":1035,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0103,"confidence_interval":0.9428,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8424,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":871,"p95_latency_ms":1065,"p99_latency_ms":1142,"time_to_first_token_ms":283,"total_runtime_ms":968,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0205,"confidence_interval":0.932,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5849,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":910,"p95_latency_ms":1112,"p99_latency_ms":1193,"time_to_first_token_ms":274,"total_runtime_ms":1011,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0243,"confidence_interval":0.9347,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0823,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":897,"p95_latency_ms":1096,"p99_latency_ms":1176,"time_to_first_token_ms":264,"total_runtime_ms":997,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.9496,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6539,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":864,"p95_latency_ms":1056,"p99_latency_ms":1133,"time_to_first_token_ms":256,"total_runtime_ms":960,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9471,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.53,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":915,"p95_latency_ms":1119,"p99_latency_ms":1200,"time_to_first_token_ms":279,"total_runtime_ms":1017,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0207,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6429,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"rewrite_constrained|high|qwen2:14b|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":871,"p95_latency_ms":1065,"p99_latency_ms":1143,"time_to_first_token_ms":256,"total_runtime_ms":968,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.014,"confidence_interval":0.9463,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7936,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":774,"p95_latency_ms":946,"p99_latency_ms":1014,"time_to_first_token_ms":202,"total_runtime_ms":860,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0197,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5282,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1030,"p99_latency_ms":1105,"time_to_first_token_ms":218,"total_runtime_ms":936,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0186,"confidence_interval":0.9347,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4427,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":808,"p95_latency_ms":987,"p99_latency_ms":1059,"time_to_first_token_ms":201,"total_runtime_ms":897,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0216,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4847,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":842,"p95_latency_ms":1029,"p99_latency_ms":1104,"time_to_first_token_ms":216,"total_runtime_ms":935,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0159,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2026,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":824,"p95_latency_ms":1007,"p99_latency_ms":1080,"time_to_first_token_ms":218,"total_runtime_ms":915,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0108,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1493,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":190,"total_runtime_ms":849,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.027,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1726,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":816,"p95_latency_ms":998,"p99_latency_ms":1070,"time_to_first_token_ms":219,"total_runtime_ms":907,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0177,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8684,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":811,"p95_latency_ms":992,"p99_latency_ms":1064,"time_to_first_token_ms":200,"total_runtime_ms":901,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0222,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3165,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":757,"p95_latency_ms":925,"p99_latency_ms":992,"time_to_first_token_ms":203,"total_runtime_ms":841,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0271,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0586,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":801,"p95_latency_ms":979,"p99_latency_ms":1050,"time_to_first_token_ms":214,"total_runtime_ms":890,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0128,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6721,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":416,"p95_latency_ms":508,"p99_latency_ms":545,"time_to_first_token_ms":155,"total_runtime_ms":462,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0114,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2928,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":380,"p95_latency_ms":465,"p99_latency_ms":499,"time_to_first_token_ms":145,"total_runtime_ms":423,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0211,"confidence_interval":0.9477,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9443,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":449,"p95_latency_ms":549,"p99_latency_ms":588,"time_to_first_token_ms":147,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0296,"confidence_interval":0.9468,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3458,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":471,"p95_latency_ms":575,"p99_latency_ms":617,"time_to_first_token_ms":143,"total_runtime_ms":523,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0185,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8634,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":480,"p95_latency_ms":587,"p99_latency_ms":629,"time_to_first_token_ms":158,"total_runtime_ms":533,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5685,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":375,"p95_latency_ms":458,"p99_latency_ms":491,"time_to_first_token_ms":165,"total_runtime_ms":416,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0277,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3686,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":155,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.019,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3368,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":394,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":174,"total_runtime_ms":437,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0249,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1271,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":417,"p95_latency_ms":510,"p99_latency_ms":547,"time_to_first_token_ms":150,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5693,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":443,"p95_latency_ms":542,"p99_latency_ms":581,"time_to_first_token_ms":147,"total_runtime_ms":493,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.017,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3324,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1118,"p95_latency_ms":1366,"p99_latency_ms":1465,"time_to_first_token_ms":249,"total_runtime_ms":1242,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0192,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1429,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1011,"p95_latency_ms":1236,"p99_latency_ms":1326,"time_to_first_token_ms":242,"total_runtime_ms":1124,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0197,"confidence_interval":0.9424,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0659,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1036,"p95_latency_ms":1266,"p99_latency_ms":1358,"time_to_first_token_ms":234,"total_runtime_ms":1151,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0104,"confidence_interval":0.9308,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6223,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1083,"p95_latency_ms":1323,"p99_latency_ms":1420,"time_to_first_token_ms":262,"total_runtime_ms":1203,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0267,"confidence_interval":0.9409,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.177,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1045,"p95_latency_ms":1277,"p99_latency_ms":1370,"time_to_first_token_ms":267,"total_runtime_ms":1161,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0292,"confidence_interval":0.9448,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.126,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1067,"p95_latency_ms":1304,"p99_latency_ms":1399,"time_to_first_token_ms":265,"total_runtime_ms":1185,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0177,"confidence_interval":0.9401,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.058,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1112,"p95_latency_ms":1359,"p99_latency_ms":1457,"time_to_first_token_ms":246,"total_runtime_ms":1235,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.016,"confidence_interval":0.9439,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.263,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1344,"time_to_first_token_ms":251,"total_runtime_ms":1139,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0145,"confidence_interval":0.9427,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.257,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1100,"p95_latency_ms":1345,"p99_latency_ms":1443,"time_to_first_token_ms":250,"total_runtime_ms":1223,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0152,"confidence_interval":0.9361,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1256,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1014,"p95_latency_ms":1239,"p99_latency_ms":1329,"time_to_first_token_ms":252,"total_runtime_ms":1127,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0223,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0892,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":531,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":213,"total_runtime_ms":590,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0158,"confidence_interval":0.9435,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.487,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":566,"p95_latency_ms":692,"p99_latency_ms":742,"time_to_first_token_ms":205,"total_runtime_ms":629,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0174,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3724,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":189,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0158,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0861,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":212,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0258,"confidence_interval":0.9472,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7502,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":538,"p95_latency_ms":657,"p99_latency_ms":705,"time_to_first_token_ms":190,"total_runtime_ms":598,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0143,"confidence_interval":0.9431,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3549,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":596,"p95_latency_ms":729,"p99_latency_ms":782,"time_to_first_token_ms":198,"total_runtime_ms":662,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0157,"confidence_interval":0.9445,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4829,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":562,"p95_latency_ms":687,"p99_latency_ms":737,"time_to_first_token_ms":217,"total_runtime_ms":624,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0125,"confidence_interval":0.9399,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5922,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":513,"p95_latency_ms":627,"p99_latency_ms":672,"time_to_first_token_ms":216,"total_runtime_ms":570,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0188,"confidence_interval":0.9423,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3627,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":206,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0171,"confidence_interval":0.9475,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2696,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":559,"p95_latency_ms":683,"p99_latency_ms":733,"time_to_first_token_ms":197,"total_runtime_ms":621,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0123,"confidence_interval":0.9419,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8161,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2089,"p95_latency_ms":2553,"p99_latency_ms":2739,"time_to_first_token_ms":435,"total_runtime_ms":2321,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9417,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2134,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2023,"p95_latency_ms":2472,"p99_latency_ms":2652,"time_to_first_token_ms":409,"total_runtime_ms":2248,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2676,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2037,"p95_latency_ms":2490,"p99_latency_ms":2671,"time_to_first_token_ms":418,"total_runtime_ms":2264,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0172,"confidence_interval":0.9301,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7058,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2124,"p95_latency_ms":2596,"p99_latency_ms":2785,"time_to_first_token_ms":432,"total_runtime_ms":2360,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.02,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7236,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2115,"p95_latency_ms":2585,"p99_latency_ms":2773,"time_to_first_token_ms":432,"total_runtime_ms":2350,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0187,"confidence_interval":0.9458,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.484,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2041,"p95_latency_ms":2495,"p99_latency_ms":2677,"time_to_first_token_ms":436,"total_runtime_ms":2268,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0105,"confidence_interval":0.9434,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4708,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2029,"p95_latency_ms":2480,"p99_latency_ms":2660,"time_to_first_token_ms":418,"total_runtime_ms":2254,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.016,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3616,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2057,"p95_latency_ms":2514,"p99_latency_ms":2697,"time_to_first_token_ms":417,"total_runtime_ms":2285,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0139,"confidence_interval":0.94,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4864,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2129,"p95_latency_ms":2603,"p99_latency_ms":2792,"time_to_first_token_ms":437,"total_runtime_ms":2366,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0156,"confidence_interval":0.9492,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.3006,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2064,"p95_latency_ms":2522,"p99_latency_ms":2706,"time_to_first_token_ms":409,"total_runtime_ms":2293,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.021,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.3109,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1038,"p95_latency_ms":1269,"p99_latency_ms":1362,"time_to_first_token_ms":364,"total_runtime_ms":1154,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0165,"confidence_interval":0.9355,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1774,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1058,"p95_latency_ms":1293,"p99_latency_ms":1387,"time_to_first_token_ms":382,"total_runtime_ms":1176,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0287,"confidence_interval":0.9378,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2232,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1051,"p95_latency_ms":1285,"p99_latency_ms":1378,"time_to_first_token_ms":395,"total_runtime_ms":1168,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0108,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5514,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1100,"p95_latency_ms":1344,"p99_latency_ms":1442,"time_to_first_token_ms":393,"total_runtime_ms":1222,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0131,"confidence_interval":0.9401,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2815,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1027,"p95_latency_ms":1256,"p99_latency_ms":1347,"time_to_first_token_ms":362,"total_runtime_ms":1142,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0108,"confidence_interval":0.9424,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7778,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1034,"p95_latency_ms":1264,"p99_latency_ms":1356,"time_to_first_token_ms":366,"total_runtime_ms":1149,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0199,"confidence_interval":0.9355,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8201,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1085,"p95_latency_ms":1326,"p99_latency_ms":1422,"time_to_first_token_ms":388,"total_runtime_ms":1205,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0116,"confidence_interval":0.9487,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1858,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1023,"p95_latency_ms":1250,"p99_latency_ms":1341,"time_to_first_token_ms":387,"total_runtime_ms":1136,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.025,"confidence_interval":0.9451,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0417,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1045,"p95_latency_ms":1277,"p99_latency_ms":1370,"time_to_first_token_ms":388,"total_runtime_ms":1161,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0264,"confidence_interval":0.9325,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7509,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1113,"p95_latency_ms":1360,"p99_latency_ms":1459,"time_to_first_token_ms":389,"total_runtime_ms":1236,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0107,"confidence_interval":0.9404,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0464,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3011,"p95_latency_ms":3680,"p99_latency_ms":3947,"time_to_first_token_ms":584,"total_runtime_ms":3345,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0242,"confidence_interval":0.945,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9413,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3059,"p95_latency_ms":3739,"p99_latency_ms":4011,"time_to_first_token_ms":578,"total_runtime_ms":3399,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0289,"confidence_interval":0.9473,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.399,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3025,"p95_latency_ms":3697,"p99_latency_ms":3966,"time_to_first_token_ms":606,"total_runtime_ms":3361,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0205,"confidence_interval":0.9303,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2712,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3022,"p95_latency_ms":3693,"p99_latency_ms":3962,"time_to_first_token_ms":603,"total_runtime_ms":3357,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0108,"confidence_interval":0.94,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5747,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3065,"p95_latency_ms":3746,"p99_latency_ms":4018,"time_to_first_token_ms":583,"total_runtime_ms":3405,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0185,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3433,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3061,"p95_latency_ms":3741,"p99_latency_ms":4013,"time_to_first_token_ms":591,"total_runtime_ms":3401,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0151,"confidence_interval":0.9448,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7799,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3089,"p95_latency_ms":3775,"p99_latency_ms":4050,"time_to_first_token_ms":594,"total_runtime_ms":3432,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0103,"confidence_interval":0.9436,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0412,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3008,"p95_latency_ms":3676,"p99_latency_ms":3944,"time_to_first_token_ms":599,"total_runtime_ms":3342,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0299,"confidence_interval":0.9321,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3306,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3076,"p95_latency_ms":3760,"p99_latency_ms":4033,"time_to_first_token_ms":586,"total_runtime_ms":3418,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0157,"confidence_interval":0.9459,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1806,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3036,"p95_latency_ms":3711,"p99_latency_ms":3981,"time_to_first_token_ms":588,"total_runtime_ms":3374,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0191,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.534,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1559,"p95_latency_ms":1906,"p99_latency_ms":2044,"time_to_first_token_ms":552,"total_runtime_ms":1732,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0229,"confidence_interval":0.9499,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.588,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1574,"p95_latency_ms":1924,"p99_latency_ms":2064,"time_to_first_token_ms":550,"total_runtime_ms":1749,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.937,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8879,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1595,"p95_latency_ms":1949,"p99_latency_ms":2091,"time_to_first_token_ms":538,"total_runtime_ms":1772,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.028,"confidence_interval":0.9321,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.583,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1549,"p95_latency_ms":1893,"p99_latency_ms":2031,"time_to_first_token_ms":552,"total_runtime_ms":1721,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0191,"confidence_interval":0.946,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5699,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1512,"p95_latency_ms":1848,"p99_latency_ms":1982,"time_to_first_token_ms":529,"total_runtime_ms":1680,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.027,"confidence_interval":0.94,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4131,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1510,"p95_latency_ms":1846,"p99_latency_ms":1980,"time_to_first_token_ms":528,"total_runtime_ms":1678,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0192,"confidence_interval":0.9414,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9546,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1512,"p95_latency_ms":1848,"p99_latency_ms":1982,"time_to_first_token_ms":535,"total_runtime_ms":1680,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0249,"confidence_interval":0.9451,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3218,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1596,"p95_latency_ms":1951,"p99_latency_ms":2093,"time_to_first_token_ms":547,"total_runtime_ms":1774,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0109,"confidence_interval":0.9447,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7656,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1567,"p95_latency_ms":1915,"p99_latency_ms":2054,"time_to_first_token_ms":534,"total_runtime_ms":1741,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0204,"confidence_interval":0.9391,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5449,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|low|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1528,"p95_latency_ms":1867,"p99_latency_ms":2003,"time_to_first_token_ms":537,"total_runtime_ms":1697,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0166,"confidence_interval":0.9344,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9117,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":190,"total_runtime_ms":664,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0261,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9578,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":570,"p95_latency_ms":696,"p99_latency_ms":747,"time_to_first_token_ms":158,"total_runtime_ms":633,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0128,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8142,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":641,"p95_latency_ms":784,"p99_latency_ms":841,"time_to_first_token_ms":168,"total_runtime_ms":713,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0247,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6744,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":641,"p95_latency_ms":783,"p99_latency_ms":840,"time_to_first_token_ms":177,"total_runtime_ms":712,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.025,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3165,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":635,"p95_latency_ms":776,"p99_latency_ms":832,"time_to_first_token_ms":159,"total_runtime_ms":705,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0267,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7793,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":604,"p95_latency_ms":738,"p99_latency_ms":792,"time_to_first_token_ms":189,"total_runtime_ms":671,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0136,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8053,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":586,"p95_latency_ms":716,"p99_latency_ms":768,"time_to_first_token_ms":164,"total_runtime_ms":651,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0222,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1747,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":607,"p95_latency_ms":742,"p99_latency_ms":795,"time_to_first_token_ms":156,"total_runtime_ms":674,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0277,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8787,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":743,"time_to_first_token_ms":174,"total_runtime_ms":630,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0221,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7512,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":660,"p95_latency_ms":807,"p99_latency_ms":865,"time_to_first_token_ms":165,"total_runtime_ms":733,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0238,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1686,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":289,"p95_latency_ms":353,"p99_latency_ms":379,"time_to_first_token_ms":112,"total_runtime_ms":321,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0144,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4137,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":357,"p99_latency_ms":383,"time_to_first_token_ms":113,"total_runtime_ms":325,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0279,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8382,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":332,"p95_latency_ms":406,"p99_latency_ms":435,"time_to_first_token_ms":132,"total_runtime_ms":369,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0129,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1929,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":291,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":125,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0102,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6419,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":295,"p95_latency_ms":360,"p99_latency_ms":386,"time_to_first_token_ms":113,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.646,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":303,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":140,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0284,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3531,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":383,"p95_latency_ms":468,"p99_latency_ms":502,"time_to_first_token_ms":145,"total_runtime_ms":425,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0114,"confidence_interval":0.9402,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.374,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":290,"p95_latency_ms":355,"p99_latency_ms":380,"time_to_first_token_ms":119,"total_runtime_ms":322,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0145,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4253,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":478,"time_to_first_token_ms":144,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.027,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.569,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":304,"p95_latency_ms":371,"p99_latency_ms":398,"time_to_first_token_ms":136,"total_runtime_ms":337,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0159,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4819,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":816,"p95_latency_ms":997,"p99_latency_ms":1070,"time_to_first_token_ms":208,"total_runtime_ms":906,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0275,"confidence_interval":0.9328,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9869,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":863,"p95_latency_ms":1055,"p99_latency_ms":1132,"time_to_first_token_ms":192,"total_runtime_ms":959,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0239,"confidence_interval":0.9386,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6296,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":212,"total_runtime_ms":895,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0264,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5998,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1030,"p99_latency_ms":1105,"time_to_first_token_ms":210,"total_runtime_ms":937,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0228,"confidence_interval":0.9451,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7497,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":840,"p95_latency_ms":1027,"p99_latency_ms":1102,"time_to_first_token_ms":193,"total_runtime_ms":934,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0206,"confidence_interval":0.9417,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5905,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":763,"p95_latency_ms":932,"p99_latency_ms":1000,"time_to_first_token_ms":196,"total_runtime_ms":848,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0159,"confidence_interval":0.9309,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0839,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":814,"p95_latency_ms":995,"p99_latency_ms":1067,"time_to_first_token_ms":193,"total_runtime_ms":904,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0113,"confidence_interval":0.9385,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3524,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":865,"p95_latency_ms":1058,"p99_latency_ms":1134,"time_to_first_token_ms":224,"total_runtime_ms":961,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0238,"confidence_interval":0.9494,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5066,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":847,"p95_latency_ms":1035,"p99_latency_ms":1110,"time_to_first_token_ms":200,"total_runtime_ms":941,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0291,"confidence_interval":0.9378,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5146,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":777,"p95_latency_ms":950,"p99_latency_ms":1019,"time_to_first_token_ms":210,"total_runtime_ms":864,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.9472,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1593,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":424,"p95_latency_ms":518,"p99_latency_ms":556,"time_to_first_token_ms":167,"total_runtime_ms":471,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0253,"confidence_interval":0.9466,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4506,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":427,"p95_latency_ms":521,"p99_latency_ms":559,"time_to_first_token_ms":160,"total_runtime_ms":474,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.022,"confidence_interval":0.9426,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4491,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":478,"p95_latency_ms":584,"p99_latency_ms":626,"time_to_first_token_ms":168,"total_runtime_ms":531,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0143,"confidence_interval":0.9364,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1607,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":426,"p95_latency_ms":521,"p99_latency_ms":558,"time_to_first_token_ms":163,"total_runtime_ms":473,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.025,"confidence_interval":0.9442,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2408,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":442,"p95_latency_ms":540,"p99_latency_ms":579,"time_to_first_token_ms":175,"total_runtime_ms":491,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0255,"confidence_interval":0.9351,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1194,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":449,"p95_latency_ms":549,"p99_latency_ms":588,"time_to_first_token_ms":177,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0284,"confidence_interval":0.9355,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7624,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":580,"p99_latency_ms":622,"time_to_first_token_ms":168,"total_runtime_ms":527,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.01,"confidence_interval":0.9381,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0186,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":438,"p95_latency_ms":536,"p99_latency_ms":574,"time_to_first_token_ms":154,"total_runtime_ms":487,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0143,"confidence_interval":0.9405,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2139,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":163,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.015,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1816,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":477,"p95_latency_ms":584,"p99_latency_ms":626,"time_to_first_token_ms":172,"total_runtime_ms":530,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0145,"confidence_interval":0.9497,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1252,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1624,"p95_latency_ms":1985,"p99_latency_ms":2130,"time_to_first_token_ms":354,"total_runtime_ms":1805,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0251,"confidence_interval":0.9368,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2502,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1578,"p95_latency_ms":1929,"p99_latency_ms":2070,"time_to_first_token_ms":320,"total_runtime_ms":1754,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0114,"confidence_interval":0.946,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.133,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1565,"p95_latency_ms":1913,"p99_latency_ms":2052,"time_to_first_token_ms":352,"total_runtime_ms":1739,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0287,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8498,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1543,"p95_latency_ms":1886,"p99_latency_ms":2023,"time_to_first_token_ms":341,"total_runtime_ms":1715,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0168,"confidence_interval":0.9328,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5755,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1604,"p95_latency_ms":1960,"p99_latency_ms":2103,"time_to_first_token_ms":352,"total_runtime_ms":1782,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0201,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6906,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1617,"p95_latency_ms":1976,"p99_latency_ms":2120,"time_to_first_token_ms":334,"total_runtime_ms":1796,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.011,"confidence_interval":0.9488,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5095,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1583,"p95_latency_ms":1935,"p99_latency_ms":2076,"time_to_first_token_ms":324,"total_runtime_ms":1759,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0176,"confidence_interval":0.9362,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7763,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1530,"p95_latency_ms":1869,"p99_latency_ms":2005,"time_to_first_token_ms":346,"total_runtime_ms":1699,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0215,"confidence_interval":0.9352,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4683,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1614,"p95_latency_ms":1973,"p99_latency_ms":2116,"time_to_first_token_ms":330,"total_runtime_ms":1793,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0115,"confidence_interval":0.9346,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6856,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1535,"p95_latency_ms":1876,"p99_latency_ms":2013,"time_to_first_token_ms":334,"total_runtime_ms":1706,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0166,"confidence_interval":0.9321,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4171,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":840,"p95_latency_ms":1027,"p99_latency_ms":1101,"time_to_first_token_ms":309,"total_runtime_ms":933,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0242,"confidence_interval":0.9403,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5719,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":859,"p95_latency_ms":1050,"p99_latency_ms":1126,"time_to_first_token_ms":283,"total_runtime_ms":954,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0235,"confidence_interval":0.9422,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8847,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":760,"p95_latency_ms":929,"p99_latency_ms":996,"time_to_first_token_ms":287,"total_runtime_ms":844,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0212,"confidence_interval":0.9334,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.753,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":844,"p95_latency_ms":1032,"p99_latency_ms":1107,"time_to_first_token_ms":277,"total_runtime_ms":938,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0272,"confidence_interval":0.94,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6171,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":839,"p95_latency_ms":1025,"p99_latency_ms":1100,"time_to_first_token_ms":277,"total_runtime_ms":932,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0209,"confidence_interval":0.9392,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4206,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":791,"p95_latency_ms":966,"p99_latency_ms":1037,"time_to_first_token_ms":300,"total_runtime_ms":878,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0245,"confidence_interval":0.9385,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1639,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":784,"p95_latency_ms":958,"p99_latency_ms":1027,"time_to_first_token_ms":308,"total_runtime_ms":871,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0258,"confidence_interval":0.9337,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3014,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":790,"p95_latency_ms":966,"p99_latency_ms":1036,"time_to_first_token_ms":307,"total_runtime_ms":878,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0169,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0523,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":820,"p95_latency_ms":1002,"p99_latency_ms":1075,"time_to_first_token_ms":297,"total_runtime_ms":911,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0238,"confidence_interval":0.9331,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1425,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":781,"p95_latency_ms":954,"p99_latency_ms":1023,"time_to_first_token_ms":287,"total_runtime_ms":867,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0193,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6093,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2272,"p95_latency_ms":2776,"p99_latency_ms":2978,"time_to_first_token_ms":463,"total_runtime_ms":2524,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0195,"confidence_interval":0.9342,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7012,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2315,"p95_latency_ms":2830,"p99_latency_ms":3035,"time_to_first_token_ms":446,"total_runtime_ms":2572,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.012,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2515,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2291,"p95_latency_ms":2800,"p99_latency_ms":3003,"time_to_first_token_ms":454,"total_runtime_ms":2545,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0292,"confidence_interval":0.943,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2262,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2272,"p95_latency_ms":2777,"p99_latency_ms":2978,"time_to_first_token_ms":464,"total_runtime_ms":2524,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0122,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1881,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2252,"p95_latency_ms":2752,"p99_latency_ms":2952,"time_to_first_token_ms":453,"total_runtime_ms":2502,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0175,"confidence_interval":0.9394,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6662,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2323,"p95_latency_ms":2839,"p99_latency_ms":3046,"time_to_first_token_ms":475,"total_runtime_ms":2581,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0135,"confidence_interval":0.9384,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8464,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2305,"p95_latency_ms":2818,"p99_latency_ms":3023,"time_to_first_token_ms":455,"total_runtime_ms":2562,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0142,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.099,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2346,"p95_latency_ms":2867,"p99_latency_ms":3075,"time_to_first_token_ms":475,"total_runtime_ms":2606,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0271,"confidence_interval":0.9462,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9656,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2289,"p95_latency_ms":2798,"p99_latency_ms":3002,"time_to_first_token_ms":473,"total_runtime_ms":2544,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0119,"confidence_interval":0.9359,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1022,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2295,"p95_latency_ms":2805,"p99_latency_ms":3009,"time_to_first_token_ms":451,"total_runtime_ms":2550,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0249,"confidence_interval":0.9472,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.187,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1168,"p95_latency_ms":1427,"p99_latency_ms":1531,"time_to_first_token_ms":404,"total_runtime_ms":1298,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0235,"confidence_interval":0.9329,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5627,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1194,"p95_latency_ms":1459,"p99_latency_ms":1565,"time_to_first_token_ms":417,"total_runtime_ms":1326,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0247,"confidence_interval":0.9439,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4923,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1168,"p95_latency_ms":1428,"p99_latency_ms":1531,"time_to_first_token_ms":428,"total_runtime_ms":1298,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0169,"confidence_interval":0.9307,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1821,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1198,"p95_latency_ms":1464,"p99_latency_ms":1570,"time_to_first_token_ms":414,"total_runtime_ms":1331,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0164,"confidence_interval":0.9338,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8307,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1213,"p95_latency_ms":1483,"p99_latency_ms":1591,"time_to_first_token_ms":415,"total_runtime_ms":1348,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0171,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7214,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1133,"p95_latency_ms":1385,"p99_latency_ms":1485,"time_to_first_token_ms":402,"total_runtime_ms":1259,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0182,"confidence_interval":0.9392,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6952,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1159,"p95_latency_ms":1417,"p99_latency_ms":1520,"time_to_first_token_ms":423,"total_runtime_ms":1288,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0183,"confidence_interval":0.94,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5156,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1145,"p95_latency_ms":1399,"p99_latency_ms":1501,"time_to_first_token_ms":399,"total_runtime_ms":1272,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0288,"confidence_interval":0.9446,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7861,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1149,"p95_latency_ms":1404,"p99_latency_ms":1506,"time_to_first_token_ms":407,"total_runtime_ms":1276,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0124,"confidence_interval":0.9457,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3042,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1178,"p95_latency_ms":1440,"p99_latency_ms":1545,"time_to_first_token_ms":429,"total_runtime_ms":1309,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0241,"confidence_interval":0.9489,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8194,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":499,"p95_latency_ms":610,"p99_latency_ms":655,"time_to_first_token_ms":145,"total_runtime_ms":555,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0175,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9089,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":469,"p95_latency_ms":573,"p99_latency_ms":615,"time_to_first_token_ms":136,"total_runtime_ms":521,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0283,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4969,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":473,"p95_latency_ms":578,"p99_latency_ms":620,"time_to_first_token_ms":134,"total_runtime_ms":526,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.029,"confidence_interval":0.9416,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4639,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":497,"p95_latency_ms":608,"p99_latency_ms":652,"time_to_first_token_ms":136,"total_runtime_ms":552,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0266,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0352,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":489,"p95_latency_ms":597,"p99_latency_ms":641,"time_to_first_token_ms":135,"total_runtime_ms":543,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0122,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.483,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":151,"total_runtime_ms":584,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0221,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4796,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":591,"time_to_first_token_ms":143,"total_runtime_ms":501,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0136,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9004,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":460,"p95_latency_ms":562,"p99_latency_ms":603,"time_to_first_token_ms":163,"total_runtime_ms":511,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.028,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5259,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":517,"p95_latency_ms":631,"p99_latency_ms":677,"time_to_first_token_ms":133,"total_runtime_ms":574,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0133,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8256,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":160,"total_runtime_ms":576,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0198,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.016,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":243,"p95_latency_ms":297,"p99_latency_ms":318,"time_to_first_token_ms":116,"total_runtime_ms":270,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0133,"confidence_interval":0.9424,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3183,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":264,"p95_latency_ms":322,"p99_latency_ms":346,"time_to_first_token_ms":91,"total_runtime_ms":293,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0163,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1219,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":240,"p95_latency_ms":293,"p99_latency_ms":314,"time_to_first_token_ms":92,"total_runtime_ms":266,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0131,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3405,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":244,"p95_latency_ms":298,"p99_latency_ms":320,"time_to_first_token_ms":99,"total_runtime_ms":271,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0251,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1032,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":237,"p95_latency_ms":290,"p99_latency_ms":311,"time_to_first_token_ms":120,"total_runtime_ms":264,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0118,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4318,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":306,"p95_latency_ms":375,"p99_latency_ms":402,"time_to_first_token_ms":115,"total_runtime_ms":341,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0276,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6946,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":232,"p95_latency_ms":284,"p99_latency_ms":305,"time_to_first_token_ms":91,"total_runtime_ms":258,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0268,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.508,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":261,"p95_latency_ms":319,"p99_latency_ms":343,"time_to_first_token_ms":109,"total_runtime_ms":290,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0268,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1317,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":301,"p95_latency_ms":368,"p99_latency_ms":394,"time_to_first_token_ms":90,"total_runtime_ms":334,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.025,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7415,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":231,"p95_latency_ms":282,"p99_latency_ms":303,"time_to_first_token_ms":92,"total_runtime_ms":256,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0258,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0346,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":621,"p95_latency_ms":759,"p99_latency_ms":814,"time_to_first_token_ms":179,"total_runtime_ms":690,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0152,"confidence_interval":0.938,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6131,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":636,"p95_latency_ms":777,"p99_latency_ms":833,"time_to_first_token_ms":166,"total_runtime_ms":706,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0271,"confidence_interval":0.9427,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2099,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":669,"p95_latency_ms":818,"p99_latency_ms":878,"time_to_first_token_ms":174,"total_runtime_ms":744,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0212,"confidence_interval":0.9336,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9446,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":167,"total_runtime_ms":741,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0163,"confidence_interval":0.9441,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9419,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":756,"time_to_first_token_ms":178,"total_runtime_ms":641,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0292,"confidence_interval":0.9459,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5663,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":661,"p95_latency_ms":807,"p99_latency_ms":866,"time_to_first_token_ms":184,"total_runtime_ms":734,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0181,"confidence_interval":0.9375,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3256,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":582,"p95_latency_ms":711,"p99_latency_ms":763,"time_to_first_token_ms":185,"total_runtime_ms":646,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0132,"confidence_interval":0.9412,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.91,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":629,"p95_latency_ms":769,"p99_latency_ms":825,"time_to_first_token_ms":171,"total_runtime_ms":699,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0289,"confidence_interval":0.9451,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5083,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":625,"p95_latency_ms":764,"p99_latency_ms":819,"time_to_first_token_ms":190,"total_runtime_ms":694,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0214,"confidence_interval":0.9497,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6168,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":611,"p95_latency_ms":747,"p99_latency_ms":801,"time_to_first_token_ms":168,"total_runtime_ms":679,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0252,"confidence_interval":0.9362,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4512,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":325,"p95_latency_ms":397,"p99_latency_ms":426,"time_to_first_token_ms":117,"total_runtime_ms":361,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0129,"confidence_interval":0.9416,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0259,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":316,"p95_latency_ms":386,"p99_latency_ms":414,"time_to_first_token_ms":130,"total_runtime_ms":351,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2438,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":357,"p95_latency_ms":437,"p99_latency_ms":469,"time_to_first_token_ms":134,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0147,"confidence_interval":0.9449,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3718,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":286,"p95_latency_ms":349,"p99_latency_ms":374,"time_to_first_token_ms":138,"total_runtime_ms":317,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0156,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8727,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":381,"p99_latency_ms":408,"time_to_first_token_ms":130,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0274,"confidence_interval":0.9315,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.337,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":478,"time_to_first_token_ms":120,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0208,"confidence_interval":0.9392,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9942,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":349,"p95_latency_ms":426,"p99_latency_ms":457,"time_to_first_token_ms":116,"total_runtime_ms":387,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0288,"confidence_interval":0.9424,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3339,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":361,"p95_latency_ms":441,"p99_latency_ms":473,"time_to_first_token_ms":145,"total_runtime_ms":401,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0153,"confidence_interval":0.9326,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5447,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":477,"time_to_first_token_ms":131,"total_runtime_ms":404,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.027,"confidence_interval":0.9345,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1344,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":291,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":126,"total_runtime_ms":323,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0118,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5311,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1173,"p95_latency_ms":1433,"p99_latency_ms":1537,"time_to_first_token_ms":265,"total_runtime_ms":1303,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9385,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3969,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1157,"p95_latency_ms":1414,"p99_latency_ms":1517,"time_to_first_token_ms":264,"total_runtime_ms":1286,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0259,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2226,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1206,"p95_latency_ms":1473,"p99_latency_ms":1581,"time_to_first_token_ms":268,"total_runtime_ms":1340,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0269,"confidence_interval":0.9335,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1209,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1205,"p95_latency_ms":1473,"p99_latency_ms":1580,"time_to_first_token_ms":260,"total_runtime_ms":1339,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.02,"confidence_interval":0.9373,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6265,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1219,"p95_latency_ms":1490,"p99_latency_ms":1599,"time_to_first_token_ms":284,"total_runtime_ms":1355,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0262,"confidence_interval":0.9413,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9713,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1208,"p95_latency_ms":1476,"p99_latency_ms":1584,"time_to_first_token_ms":268,"total_runtime_ms":1342,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0292,"confidence_interval":0.9477,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6691,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1241,"p95_latency_ms":1516,"p99_latency_ms":1627,"time_to_first_token_ms":259,"total_runtime_ms":1378,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0187,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.989,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1228,"p95_latency_ms":1501,"p99_latency_ms":1610,"time_to_first_token_ms":282,"total_runtime_ms":1364,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0266,"confidence_interval":0.9352,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6929,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1231,"p95_latency_ms":1505,"p99_latency_ms":1614,"time_to_first_token_ms":256,"total_runtime_ms":1368,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0133,"confidence_interval":0.947,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7181,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1240,"p95_latency_ms":1515,"p99_latency_ms":1626,"time_to_first_token_ms":279,"total_runtime_ms":1378,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0143,"confidence_interval":0.9373,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8196,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":573,"p95_latency_ms":700,"p99_latency_ms":751,"time_to_first_token_ms":235,"total_runtime_ms":637,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1958,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":625,"p95_latency_ms":764,"p99_latency_ms":819,"time_to_first_token_ms":223,"total_runtime_ms":694,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0193,"confidence_interval":0.9358,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6393,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":642,"p95_latency_ms":785,"p99_latency_ms":842,"time_to_first_token_ms":242,"total_runtime_ms":713,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.019,"confidence_interval":0.9374,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9776,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":602,"p95_latency_ms":736,"p99_latency_ms":789,"time_to_first_token_ms":217,"total_runtime_ms":669,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0156,"confidence_interval":0.9393,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2294,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":676,"p95_latency_ms":826,"p99_latency_ms":886,"time_to_first_token_ms":239,"total_runtime_ms":751,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0127,"confidence_interval":0.9347,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3752,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":226,"total_runtime_ms":635,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0193,"confidence_interval":0.9318,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8473,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":867,"time_to_first_token_ms":216,"total_runtime_ms":735,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0211,"confidence_interval":0.9391,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0993,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":578,"p95_latency_ms":706,"p99_latency_ms":758,"time_to_first_token_ms":231,"total_runtime_ms":642,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0196,"confidence_interval":0.9343,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2919,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":677,"p95_latency_ms":827,"p99_latency_ms":887,"time_to_first_token_ms":212,"total_runtime_ms":752,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0121,"confidence_interval":0.9428,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4333,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":658,"p95_latency_ms":804,"p99_latency_ms":863,"time_to_first_token_ms":214,"total_runtime_ms":731,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0277,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6768,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1718,"p95_latency_ms":2100,"p99_latency_ms":2253,"time_to_first_token_ms":355,"total_runtime_ms":1909,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0112,"confidence_interval":0.9392,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0082,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1752,"p95_latency_ms":2141,"p99_latency_ms":2297,"time_to_first_token_ms":364,"total_runtime_ms":1947,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0279,"confidence_interval":0.9384,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5241,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1707,"p95_latency_ms":2087,"p99_latency_ms":2239,"time_to_first_token_ms":353,"total_runtime_ms":1897,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0201,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9737,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1689,"p95_latency_ms":2065,"p99_latency_ms":2215,"time_to_first_token_ms":382,"total_runtime_ms":1877,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0228,"confidence_interval":0.9471,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7702,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1692,"p95_latency_ms":2068,"p99_latency_ms":2218,"time_to_first_token_ms":349,"total_runtime_ms":1880,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0179,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6018,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1784,"p95_latency_ms":2180,"p99_latency_ms":2339,"time_to_first_token_ms":370,"total_runtime_ms":1982,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0257,"confidence_interval":0.9483,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2119,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1782,"p95_latency_ms":2178,"p99_latency_ms":2336,"time_to_first_token_ms":354,"total_runtime_ms":1980,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0122,"confidence_interval":0.9466,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.613,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1768,"p95_latency_ms":2161,"p99_latency_ms":2318,"time_to_first_token_ms":373,"total_runtime_ms":1964,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0173,"confidence_interval":0.9395,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8326,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1738,"p95_latency_ms":2125,"p99_latency_ms":2279,"time_to_first_token_ms":359,"total_runtime_ms":1932,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0186,"confidence_interval":0.9496,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.954,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1706,"p95_latency_ms":2085,"p99_latency_ms":2236,"time_to_first_token_ms":379,"total_runtime_ms":1895,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0234,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8587,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":931,"p95_latency_ms":1137,"p99_latency_ms":1220,"time_to_first_token_ms":317,"total_runtime_ms":1034,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0193,"confidence_interval":0.9423,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9459,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":905,"p95_latency_ms":1106,"p99_latency_ms":1187,"time_to_first_token_ms":309,"total_runtime_ms":1006,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0254,"confidence_interval":0.9419,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2819,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":904,"p95_latency_ms":1105,"p99_latency_ms":1185,"time_to_first_token_ms":314,"total_runtime_ms":1004,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0241,"confidence_interval":0.9466,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7437,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":908,"p95_latency_ms":1110,"p99_latency_ms":1191,"time_to_first_token_ms":328,"total_runtime_ms":1009,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0166,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5361,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":888,"p95_latency_ms":1085,"p99_latency_ms":1164,"time_to_first_token_ms":311,"total_runtime_ms":987,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9387,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0105,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":925,"p95_latency_ms":1131,"p99_latency_ms":1213,"time_to_first_token_ms":308,"total_runtime_ms":1028,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0141,"confidence_interval":0.949,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8835,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":945,"p95_latency_ms":1155,"p99_latency_ms":1239,"time_to_first_token_ms":327,"total_runtime_ms":1050,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.026,"confidence_interval":0.9327,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4967,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":945,"p95_latency_ms":1156,"p99_latency_ms":1240,"time_to_first_token_ms":306,"total_runtime_ms":1050,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0136,"confidence_interval":0.9341,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9128,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":858,"p95_latency_ms":1048,"p99_latency_ms":1124,"time_to_first_token_ms":333,"total_runtime_ms":953,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0246,"confidence_interval":0.9446,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5947,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":879,"p95_latency_ms":1074,"p99_latency_ms":1152,"time_to_first_token_ms":332,"total_runtime_ms":977,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0127,"confidence_interval":0.9338,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9009,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":495,"time_to_first_token_ms":135,"total_runtime_ms":420,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0261,"confidence_interval":0.944,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4492,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":415,"p95_latency_ms":507,"p99_latency_ms":544,"time_to_first_token_ms":147,"total_runtime_ms":461,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0232,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3907,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":429,"p95_latency_ms":524,"p99_latency_ms":562,"time_to_first_token_ms":140,"total_runtime_ms":476,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0267,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9957,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":373,"p95_latency_ms":456,"p99_latency_ms":489,"time_to_first_token_ms":137,"total_runtime_ms":415,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0137,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2937,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":382,"p95_latency_ms":467,"p99_latency_ms":501,"time_to_first_token_ms":153,"total_runtime_ms":424,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0247,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3041,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":356,"p95_latency_ms":435,"p99_latency_ms":467,"time_to_first_token_ms":143,"total_runtime_ms":396,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0196,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8472,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":131,"total_runtime_ms":504,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0153,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7858,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":457,"p95_latency_ms":559,"p99_latency_ms":599,"time_to_first_token_ms":129,"total_runtime_ms":508,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4205,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":433,"p95_latency_ms":529,"p99_latency_ms":568,"time_to_first_token_ms":135,"total_runtime_ms":481,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0153,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3909,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":390,"p95_latency_ms":477,"p99_latency_ms":512,"time_to_first_token_ms":128,"total_runtime_ms":434,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0281,"confidence_interval":0.9477,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6583,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":215,"p95_latency_ms":263,"p99_latency_ms":282,"time_to_first_token_ms":102,"total_runtime_ms":239,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0207,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3152,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":261,"p95_latency_ms":320,"p99_latency_ms":343,"time_to_first_token_ms":77,"total_runtime_ms":290,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0175,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4459,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":245,"p95_latency_ms":299,"p99_latency_ms":321,"time_to_first_token_ms":81,"total_runtime_ms":272,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0219,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5508,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":205,"p95_latency_ms":251,"p99_latency_ms":269,"time_to_first_token_ms":97,"total_runtime_ms":228,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0269,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6814,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":245,"p95_latency_ms":299,"p99_latency_ms":321,"time_to_first_token_ms":100,"total_runtime_ms":272,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0111,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4688,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":199,"p95_latency_ms":243,"p99_latency_ms":261,"time_to_first_token_ms":101,"total_runtime_ms":221,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0262,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3738,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":267,"p95_latency_ms":327,"p99_latency_ms":351,"time_to_first_token_ms":95,"total_runtime_ms":297,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0104,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1858,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":267,"p95_latency_ms":326,"p99_latency_ms":350,"time_to_first_token_ms":102,"total_runtime_ms":297,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0232,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2703,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":264,"p95_latency_ms":323,"p99_latency_ms":347,"time_to_first_token_ms":106,"total_runtime_ms":294,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1243,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":248,"p95_latency_ms":304,"p99_latency_ms":326,"time_to_first_token_ms":95,"total_runtime_ms":276,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8178,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":501,"p95_latency_ms":612,"p99_latency_ms":656,"time_to_first_token_ms":166,"total_runtime_ms":556,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0221,"confidence_interval":0.9484,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2316,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":543,"p95_latency_ms":664,"p99_latency_ms":712,"time_to_first_token_ms":153,"total_runtime_ms":603,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0165,"confidence_interval":0.9372,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4623,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":528,"p95_latency_ms":646,"p99_latency_ms":693,"time_to_first_token_ms":176,"total_runtime_ms":587,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0269,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9407,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":484,"p95_latency_ms":592,"p99_latency_ms":635,"time_to_first_token_ms":173,"total_runtime_ms":538,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0126,"confidence_interval":0.9401,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3888,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":582,"p95_latency_ms":711,"p99_latency_ms":763,"time_to_first_token_ms":169,"total_runtime_ms":646,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0207,"confidence_interval":0.941,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.017,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":475,"p95_latency_ms":581,"p99_latency_ms":623,"time_to_first_token_ms":155,"total_runtime_ms":528,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0255,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3276,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":534,"p95_latency_ms":652,"p99_latency_ms":699,"time_to_first_token_ms":158,"total_runtime_ms":593,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0116,"confidence_interval":0.9315,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0731,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":519,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":150,"total_runtime_ms":576,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0185,"confidence_interval":0.9308,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8961,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":485,"p95_latency_ms":593,"p99_latency_ms":636,"time_to_first_token_ms":169,"total_runtime_ms":539,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.01,"confidence_interval":0.9404,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9556,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":174,"total_runtime_ms":536,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0255,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4343,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":260,"p95_latency_ms":318,"p99_latency_ms":341,"time_to_first_token_ms":125,"total_runtime_ms":289,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0276,"confidence_interval":0.9316,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2098,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":319,"p95_latency_ms":390,"p99_latency_ms":418,"time_to_first_token_ms":101,"total_runtime_ms":354,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0295,"confidence_interval":0.9424,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4514,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":259,"p95_latency_ms":316,"p99_latency_ms":339,"time_to_first_token_ms":117,"total_runtime_ms":288,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0148,"confidence_interval":0.944,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.839,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":279,"p95_latency_ms":341,"p99_latency_ms":366,"time_to_first_token_ms":130,"total_runtime_ms":310,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0206,"confidence_interval":0.9442,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0597,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":246,"p95_latency_ms":301,"p99_latency_ms":323,"time_to_first_token_ms":105,"total_runtime_ms":274,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.026,"confidence_interval":0.9467,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4981,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":326,"p95_latency_ms":398,"p99_latency_ms":427,"time_to_first_token_ms":111,"total_runtime_ms":362,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0183,"confidence_interval":0.9456,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7536,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":320,"p95_latency_ms":391,"p99_latency_ms":419,"time_to_first_token_ms":124,"total_runtime_ms":355,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0211,"confidence_interval":0.9317,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1157,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":407,"time_to_first_token_ms":120,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0217,"confidence_interval":0.9476,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5901,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":333,"p95_latency_ms":407,"p99_latency_ms":437,"time_to_first_token_ms":101,"total_runtime_ms":370,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0298,"confidence_interval":0.9385,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2652,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":268,"p95_latency_ms":327,"p99_latency_ms":351,"time_to_first_token_ms":103,"total_runtime_ms":297,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9356,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5729,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":984,"p95_latency_ms":1203,"p99_latency_ms":1290,"time_to_first_token_ms":242,"total_runtime_ms":1093,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0135,"confidence_interval":0.9462,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1959,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1027,"p95_latency_ms":1255,"p99_latency_ms":1346,"time_to_first_token_ms":239,"total_runtime_ms":1141,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0262,"confidence_interval":0.9496,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4285,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":967,"p95_latency_ms":1182,"p99_latency_ms":1267,"time_to_first_token_ms":231,"total_runtime_ms":1074,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0203,"confidence_interval":0.9325,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7264,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1053,"p95_latency_ms":1287,"p99_latency_ms":1380,"time_to_first_token_ms":227,"total_runtime_ms":1170,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0201,"confidence_interval":0.9362,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8557,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1042,"p95_latency_ms":1273,"p99_latency_ms":1366,"time_to_first_token_ms":233,"total_runtime_ms":1157,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0148,"confidence_interval":0.9473,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1182,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":953,"p95_latency_ms":1165,"p99_latency_ms":1249,"time_to_first_token_ms":256,"total_runtime_ms":1059,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0296,"confidence_interval":0.9426,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4697,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":999,"p95_latency_ms":1221,"p99_latency_ms":1310,"time_to_first_token_ms":231,"total_runtime_ms":1110,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0117,"confidence_interval":0.9481,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8901,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":949,"p95_latency_ms":1160,"p99_latency_ms":1244,"time_to_first_token_ms":225,"total_runtime_ms":1054,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0119,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5987,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1036,"p95_latency_ms":1266,"p99_latency_ms":1358,"time_to_first_token_ms":250,"total_runtime_ms":1151,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0188,"confidence_interval":0.9414,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0746,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1035,"p95_latency_ms":1265,"p99_latency_ms":1357,"time_to_first_token_ms":249,"total_runtime_ms":1150,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9395,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.018,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":489,"p95_latency_ms":598,"p99_latency_ms":642,"time_to_first_token_ms":204,"total_runtime_ms":544,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0125,"confidence_interval":0.9438,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4489,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":574,"p95_latency_ms":702,"p99_latency_ms":753,"time_to_first_token_ms":195,"total_runtime_ms":638,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.9371,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.342,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":477,"p95_latency_ms":583,"p99_latency_ms":626,"time_to_first_token_ms":204,"total_runtime_ms":530,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0264,"confidence_interval":0.9408,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9795,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":560,"p95_latency_ms":685,"p99_latency_ms":735,"time_to_first_token_ms":210,"total_runtime_ms":623,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0226,"confidence_interval":0.9393,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3473,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":503,"p95_latency_ms":615,"p99_latency_ms":659,"time_to_first_token_ms":200,"total_runtime_ms":559,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0106,"confidence_interval":0.9419,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7012,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":514,"p95_latency_ms":628,"p99_latency_ms":673,"time_to_first_token_ms":191,"total_runtime_ms":571,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0238,"confidence_interval":0.9366,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8386,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":203,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0161,"confidence_interval":0.9433,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9709,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":502,"p95_latency_ms":613,"p99_latency_ms":658,"time_to_first_token_ms":200,"total_runtime_ms":557,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0208,"confidence_interval":0.9472,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.749,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":501,"p95_latency_ms":612,"p99_latency_ms":657,"time_to_first_token_ms":191,"total_runtime_ms":556,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1818,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":190,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0285,"confidence_interval":0.9381,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9175,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1439,"p95_latency_ms":1759,"p99_latency_ms":1887,"time_to_first_token_ms":319,"total_runtime_ms":1599,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0129,"confidence_interval":0.9446,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6935,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1431,"p95_latency_ms":1749,"p99_latency_ms":1876,"time_to_first_token_ms":318,"total_runtime_ms":1590,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0226,"confidence_interval":0.9345,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1823,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1492,"p95_latency_ms":1824,"p99_latency_ms":1957,"time_to_first_token_ms":324,"total_runtime_ms":1658,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0105,"confidence_interval":0.9311,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2054,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1414,"p95_latency_ms":1729,"p99_latency_ms":1855,"time_to_first_token_ms":314,"total_runtime_ms":1572,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0211,"confidence_interval":0.9301,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8862,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1438,"p95_latency_ms":1757,"p99_latency_ms":1885,"time_to_first_token_ms":317,"total_runtime_ms":1597,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0204,"confidence_interval":0.9361,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7396,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1508,"p95_latency_ms":1843,"p99_latency_ms":1977,"time_to_first_token_ms":332,"total_runtime_ms":1675,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0296,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0077,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1438,"p95_latency_ms":1757,"p99_latency_ms":1885,"time_to_first_token_ms":301,"total_runtime_ms":1598,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0185,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5964,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1491,"p95_latency_ms":1822,"p99_latency_ms":1955,"time_to_first_token_ms":332,"total_runtime_ms":1657,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0133,"confidence_interval":0.9318,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9053,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1489,"p95_latency_ms":1820,"p99_latency_ms":1952,"time_to_first_token_ms":315,"total_runtime_ms":1654,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0102,"confidence_interval":0.9396,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5341,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1452,"p95_latency_ms":1775,"p99_latency_ms":1904,"time_to_first_token_ms":300,"total_runtime_ms":1614,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0256,"confidence_interval":0.9301,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1508,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":799,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":282,"total_runtime_ms":887,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0251,"confidence_interval":0.9423,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5091,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":755,"p95_latency_ms":922,"p99_latency_ms":989,"time_to_first_token_ms":282,"total_runtime_ms":839,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0224,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9347,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":777,"p95_latency_ms":949,"p99_latency_ms":1018,"time_to_first_token_ms":274,"total_runtime_ms":863,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0274,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8554,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":771,"p95_latency_ms":943,"p99_latency_ms":1011,"time_to_first_token_ms":277,"total_runtime_ms":857,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0251,"confidence_interval":0.9379,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3099,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":808,"p95_latency_ms":987,"p99_latency_ms":1059,"time_to_first_token_ms":285,"total_runtime_ms":897,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0132,"confidence_interval":0.9483,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4332,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":755,"p95_latency_ms":923,"p99_latency_ms":990,"time_to_first_token_ms":260,"total_runtime_ms":839,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0251,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7589,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":734,"p95_latency_ms":897,"p99_latency_ms":963,"time_to_first_token_ms":281,"total_runtime_ms":816,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0153,"confidence_interval":0.9439,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3561,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":705,"p95_latency_ms":861,"p99_latency_ms":924,"time_to_first_token_ms":286,"total_runtime_ms":783,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0124,"confidence_interval":0.9339,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7635,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":768,"p95_latency_ms":939,"p99_latency_ms":1007,"time_to_first_token_ms":288,"total_runtime_ms":853,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.022,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3368,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_binary|high|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":259,"total_runtime_ms":849,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0199,"confidence_interval":0.9391,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9832,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":905,"p95_latency_ms":1106,"p99_latency_ms":1187,"time_to_first_token_ms":203,"total_runtime_ms":1006,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0182,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1247,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":895,"p95_latency_ms":1093,"p99_latency_ms":1173,"time_to_first_token_ms":223,"total_runtime_ms":994,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9376,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0645,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":905,"p95_latency_ms":1106,"p99_latency_ms":1187,"time_to_first_token_ms":196,"total_runtime_ms":1006,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0287,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8856,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":892,"p95_latency_ms":1091,"p99_latency_ms":1170,"time_to_first_token_ms":220,"total_runtime_ms":991,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0287,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0647,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":903,"p95_latency_ms":1103,"p99_latency_ms":1184,"time_to_first_token_ms":205,"total_runtime_ms":1003,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0275,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0894,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":858,"p95_latency_ms":1049,"p99_latency_ms":1125,"time_to_first_token_ms":214,"total_runtime_ms":954,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0168,"confidence_interval":0.9361,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2056,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":884,"p95_latency_ms":1080,"p99_latency_ms":1159,"time_to_first_token_ms":190,"total_runtime_ms":982,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0293,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8018,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":213,"total_runtime_ms":937,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0101,"confidence_interval":0.95,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2073,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":897,"p95_latency_ms":1096,"p99_latency_ms":1176,"time_to_first_token_ms":213,"total_runtime_ms":997,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0286,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9947,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":867,"p95_latency_ms":1059,"p99_latency_ms":1136,"time_to_first_token_ms":193,"total_runtime_ms":963,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0268,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8581,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":483,"p95_latency_ms":590,"p99_latency_ms":633,"time_to_first_token_ms":151,"total_runtime_ms":536,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7477,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":465,"p95_latency_ms":569,"p99_latency_ms":610,"time_to_first_token_ms":174,"total_runtime_ms":517,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9388,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9147,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":165,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0256,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1969,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":483,"p95_latency_ms":590,"p99_latency_ms":633,"time_to_first_token_ms":162,"total_runtime_ms":536,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0248,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7768,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":143,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0161,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9867,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":552,"p95_latency_ms":675,"p99_latency_ms":724,"time_to_first_token_ms":163,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0183,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7698,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":462,"p95_latency_ms":564,"p99_latency_ms":605,"time_to_first_token_ms":158,"total_runtime_ms":513,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0138,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6241,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":166,"total_runtime_ms":591,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0274,"confidence_interval":0.9414,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2187,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":471,"p95_latency_ms":575,"p99_latency_ms":617,"time_to_first_token_ms":147,"total_runtime_ms":523,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0218,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8148,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":580,"p99_latency_ms":622,"time_to_first_token_ms":177,"total_runtime_ms":527,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0188,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.431,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1212,"p95_latency_ms":1481,"p99_latency_ms":1589,"time_to_first_token_ms":261,"total_runtime_ms":1346,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0289,"confidence_interval":0.9355,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6856,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1123,"p95_latency_ms":1372,"p99_latency_ms":1472,"time_to_first_token_ms":245,"total_runtime_ms":1248,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0257,"confidence_interval":0.9404,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0075,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1191,"p95_latency_ms":1455,"p99_latency_ms":1561,"time_to_first_token_ms":265,"total_runtime_ms":1323,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0262,"confidence_interval":0.9332,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5081,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1139,"p95_latency_ms":1393,"p99_latency_ms":1494,"time_to_first_token_ms":260,"total_runtime_ms":1266,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0214,"confidence_interval":0.9344,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9007,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1122,"p95_latency_ms":1372,"p99_latency_ms":1472,"time_to_first_token_ms":267,"total_runtime_ms":1247,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0149,"confidence_interval":0.9337,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9636,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1121,"p95_latency_ms":1371,"p99_latency_ms":1470,"time_to_first_token_ms":235,"total_runtime_ms":1246,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0151,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3048,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1162,"p95_latency_ms":1420,"p99_latency_ms":1523,"time_to_first_token_ms":250,"total_runtime_ms":1291,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.018,"confidence_interval":0.9381,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8112,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1126,"p95_latency_ms":1376,"p99_latency_ms":1476,"time_to_first_token_ms":240,"total_runtime_ms":1251,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0271,"confidence_interval":0.9358,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2641,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1220,"p95_latency_ms":1491,"p99_latency_ms":1599,"time_to_first_token_ms":256,"total_runtime_ms":1355,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0125,"confidence_interval":0.9405,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.656,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1142,"p95_latency_ms":1396,"p99_latency_ms":1498,"time_to_first_token_ms":266,"total_runtime_ms":1269,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0139,"confidence_interval":0.9327,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9506,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":931,"time_to_first_token_ms":219,"total_runtime_ms":789,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.029,"confidence_interval":0.9463,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4682,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":712,"p95_latency_ms":870,"p99_latency_ms":933,"time_to_first_token_ms":218,"total_runtime_ms":791,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0177,"confidence_interval":0.9493,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9294,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":671,"p95_latency_ms":820,"p99_latency_ms":880,"time_to_first_token_ms":201,"total_runtime_ms":745,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0205,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6227,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":632,"p95_latency_ms":772,"p99_latency_ms":828,"time_to_first_token_ms":217,"total_runtime_ms":702,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0244,"confidence_interval":0.9453,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3848,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":696,"p95_latency_ms":850,"p99_latency_ms":912,"time_to_first_token_ms":190,"total_runtime_ms":773,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0221,"confidence_interval":0.9331,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6579,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":616,"p95_latency_ms":753,"p99_latency_ms":808,"time_to_first_token_ms":200,"total_runtime_ms":685,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9448,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0757,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":668,"p95_latency_ms":817,"p99_latency_ms":876,"time_to_first_token_ms":218,"total_runtime_ms":742,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0296,"confidence_interval":0.9402,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7694,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":641,"p95_latency_ms":784,"p99_latency_ms":841,"time_to_first_token_ms":195,"total_runtime_ms":713,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0201,"confidence_interval":0.932,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.948,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":206,"total_runtime_ms":741,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0218,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2382,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":671,"p95_latency_ms":820,"p99_latency_ms":880,"time_to_first_token_ms":201,"total_runtime_ms":745,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0169,"confidence_interval":0.9466,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6662,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2274,"p95_latency_ms":2780,"p99_latency_ms":2982,"time_to_first_token_ms":418,"total_runtime_ms":2527,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0224,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7146,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2294,"p95_latency_ms":2804,"p99_latency_ms":3008,"time_to_first_token_ms":439,"total_runtime_ms":2549,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0264,"confidence_interval":0.9414,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3576,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2288,"p95_latency_ms":2796,"p99_latency_ms":3000,"time_to_first_token_ms":413,"total_runtime_ms":2542,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0282,"confidence_interval":0.9302,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4997,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2277,"p95_latency_ms":2784,"p99_latency_ms":2986,"time_to_first_token_ms":425,"total_runtime_ms":2530,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0229,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5858,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2324,"p95_latency_ms":2840,"p99_latency_ms":3046,"time_to_first_token_ms":439,"total_runtime_ms":2582,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0174,"confidence_interval":0.9416,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5131,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2248,"p95_latency_ms":2748,"p99_latency_ms":2948,"time_to_first_token_ms":428,"total_runtime_ms":2498,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0168,"confidence_interval":0.9463,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7531,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2323,"p95_latency_ms":2839,"p99_latency_ms":3046,"time_to_first_token_ms":437,"total_runtime_ms":2581,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0211,"confidence_interval":0.9353,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3953,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2291,"p95_latency_ms":2801,"p99_latency_ms":3004,"time_to_first_token_ms":421,"total_runtime_ms":2546,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0139,"confidence_interval":0.9379,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.578,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2314,"p95_latency_ms":2829,"p99_latency_ms":3035,"time_to_first_token_ms":419,"total_runtime_ms":2572,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0294,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4161,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2329,"p95_latency_ms":2847,"p99_latency_ms":3054,"time_to_first_token_ms":434,"total_runtime_ms":2588,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0277,"confidence_interval":0.9417,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0784,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1270,"p95_latency_ms":1552,"p99_latency_ms":1665,"time_to_first_token_ms":382,"total_runtime_ms":1411,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0248,"confidence_interval":0.9414,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0696,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1307,"p95_latency_ms":1597,"p99_latency_ms":1714,"time_to_first_token_ms":380,"total_runtime_ms":1452,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.018,"confidence_interval":0.9376,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6955,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1231,"p95_latency_ms":1504,"p99_latency_ms":1614,"time_to_first_token_ms":393,"total_runtime_ms":1368,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0207,"confidence_interval":0.9412,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8208,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1222,"p95_latency_ms":1494,"p99_latency_ms":1603,"time_to_first_token_ms":386,"total_runtime_ms":1358,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.02,"confidence_interval":0.933,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0737,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1223,"p95_latency_ms":1495,"p99_latency_ms":1604,"time_to_first_token_ms":384,"total_runtime_ms":1359,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0222,"confidence_interval":0.9429,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5703,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1277,"p95_latency_ms":1560,"p99_latency_ms":1674,"time_to_first_token_ms":363,"total_runtime_ms":1418,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.014,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3242,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1255,"p95_latency_ms":1534,"p99_latency_ms":1646,"time_to_first_token_ms":369,"total_runtime_ms":1395,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0273,"confidence_interval":0.9481,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1617,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1260,"p95_latency_ms":1540,"p99_latency_ms":1652,"time_to_first_token_ms":368,"total_runtime_ms":1400,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0115,"confidence_interval":0.9408,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9026,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1300,"p95_latency_ms":1589,"p99_latency_ms":1705,"time_to_first_token_ms":386,"total_runtime_ms":1445,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0247,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2874,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1283,"p95_latency_ms":1568,"p99_latency_ms":1682,"time_to_first_token_ms":373,"total_runtime_ms":1425,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.022,"confidence_interval":0.9326,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9227,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3372,"p95_latency_ms":4122,"p99_latency_ms":4422,"time_to_first_token_ms":593,"total_runtime_ms":3747,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0107,"confidence_interval":0.9463,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8971,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3387,"p95_latency_ms":4139,"p99_latency_ms":4440,"time_to_first_token_ms":592,"total_runtime_ms":3763,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0118,"confidence_interval":0.9313,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6355,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3363,"p95_latency_ms":4110,"p99_latency_ms":4409,"time_to_first_token_ms":606,"total_runtime_ms":3737,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0198,"confidence_interval":0.9425,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.116,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3307,"p95_latency_ms":4041,"p99_latency_ms":4335,"time_to_first_token_ms":595,"total_runtime_ms":3674,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0121,"confidence_interval":0.9466,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6628,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3338,"p95_latency_ms":4080,"p99_latency_ms":4377,"time_to_first_token_ms":596,"total_runtime_ms":3709,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0257,"confidence_interval":0.9336,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7023,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3398,"p95_latency_ms":4153,"p99_latency_ms":4455,"time_to_first_token_ms":604,"total_runtime_ms":3776,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0156,"confidence_interval":0.9368,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.097,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3365,"p95_latency_ms":4113,"p99_latency_ms":4412,"time_to_first_token_ms":607,"total_runtime_ms":3739,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0293,"confidence_interval":0.9433,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1734,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3325,"p95_latency_ms":4064,"p99_latency_ms":4359,"time_to_first_token_ms":573,"total_runtime_ms":3694,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0234,"confidence_interval":0.9404,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1558,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3353,"p95_latency_ms":4099,"p99_latency_ms":4397,"time_to_first_token_ms":592,"total_runtime_ms":3726,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0119,"confidence_interval":0.9353,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0446,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3397,"p95_latency_ms":4152,"p99_latency_ms":4454,"time_to_first_token_ms":590,"total_runtime_ms":3775,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.019,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0641,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1895,"p95_latency_ms":2316,"p99_latency_ms":2485,"time_to_first_token_ms":561,"total_runtime_ms":2106,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0172,"confidence_interval":0.9346,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4764,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1807,"p95_latency_ms":2209,"p99_latency_ms":2370,"time_to_first_token_ms":561,"total_runtime_ms":2008,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0171,"confidence_interval":0.948,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7576,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1806,"p95_latency_ms":2207,"p99_latency_ms":2368,"time_to_first_token_ms":556,"total_runtime_ms":2007,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0212,"confidence_interval":0.9433,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7269,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1815,"p95_latency_ms":2218,"p99_latency_ms":2379,"time_to_first_token_ms":542,"total_runtime_ms":2016,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0278,"confidence_interval":0.9424,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8833,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1822,"p95_latency_ms":2227,"p99_latency_ms":2389,"time_to_first_token_ms":561,"total_runtime_ms":2025,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.026,"confidence_interval":0.9341,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7188,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1905,"p95_latency_ms":2328,"p99_latency_ms":2498,"time_to_first_token_ms":533,"total_runtime_ms":2117,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0268,"confidence_interval":0.9436,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5667,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1837,"p95_latency_ms":2246,"p99_latency_ms":2409,"time_to_first_token_ms":553,"total_runtime_ms":2041,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0202,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0387,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1836,"p95_latency_ms":2244,"p99_latency_ms":2407,"time_to_first_token_ms":558,"total_runtime_ms":2040,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0247,"confidence_interval":0.9388,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2514,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1849,"p95_latency_ms":2260,"p99_latency_ms":2425,"time_to_first_token_ms":533,"total_runtime_ms":2055,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0153,"confidence_interval":0.9393,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9421,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|low|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1848,"p95_latency_ms":2259,"p99_latency_ms":2423,"time_to_first_token_ms":554,"total_runtime_ms":2053,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.019,"confidence_interval":0.9377,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2116,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":673,"p95_latency_ms":823,"p99_latency_ms":883,"time_to_first_token_ms":188,"total_runtime_ms":748,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0189,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2083,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":625,"p95_latency_ms":764,"p99_latency_ms":819,"time_to_first_token_ms":157,"total_runtime_ms":694,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0239,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3029,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":698,"p95_latency_ms":853,"p99_latency_ms":915,"time_to_first_token_ms":175,"total_runtime_ms":775,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0213,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2232,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":866,"time_to_first_token_ms":187,"total_runtime_ms":734,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.012,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9657,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":645,"p95_latency_ms":788,"p99_latency_ms":846,"time_to_first_token_ms":190,"total_runtime_ms":717,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.015,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4889,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":721,"p95_latency_ms":881,"p99_latency_ms":946,"time_to_first_token_ms":179,"total_runtime_ms":801,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0184,"confidence_interval":0.9414,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6131,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":629,"p95_latency_ms":769,"p99_latency_ms":825,"time_to_first_token_ms":178,"total_runtime_ms":699,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0294,"confidence_interval":0.9442,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.435,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":655,"p95_latency_ms":801,"p99_latency_ms":859,"time_to_first_token_ms":161,"total_runtime_ms":728,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0198,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8696,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":668,"p95_latency_ms":816,"p99_latency_ms":875,"time_to_first_token_ms":169,"total_runtime_ms":742,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.012,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9846,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":622,"p95_latency_ms":760,"p99_latency_ms":815,"time_to_first_token_ms":177,"total_runtime_ms":691,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0249,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8808,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":343,"p95_latency_ms":419,"p99_latency_ms":450,"time_to_first_token_ms":134,"total_runtime_ms":381,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0141,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2723,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":373,"p95_latency_ms":456,"p99_latency_ms":489,"time_to_first_token_ms":131,"total_runtime_ms":414,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.029,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7029,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":513,"time_to_first_token_ms":127,"total_runtime_ms":434,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0219,"confidence_interval":0.9362,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1279,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":381,"p95_latency_ms":466,"p99_latency_ms":499,"time_to_first_token_ms":121,"total_runtime_ms":423,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0274,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6258,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":342,"p95_latency_ms":418,"p99_latency_ms":448,"time_to_first_token_ms":141,"total_runtime_ms":380,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0116,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5144,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":427,"p95_latency_ms":522,"p99_latency_ms":560,"time_to_first_token_ms":128,"total_runtime_ms":475,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0235,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1726,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":384,"p95_latency_ms":469,"p99_latency_ms":503,"time_to_first_token_ms":123,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0193,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8303,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":434,"p95_latency_ms":530,"p99_latency_ms":569,"time_to_first_token_ms":117,"total_runtime_ms":482,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0279,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9333,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":144,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0213,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0963,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":402,"p95_latency_ms":491,"p99_latency_ms":527,"time_to_first_token_ms":117,"total_runtime_ms":446,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.028,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5049,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":919,"p95_latency_ms":1124,"p99_latency_ms":1205,"time_to_first_token_ms":214,"total_runtime_ms":1021,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0225,"confidence_interval":0.9362,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.419,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":880,"p95_latency_ms":1075,"p99_latency_ms":1154,"time_to_first_token_ms":214,"total_runtime_ms":978,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0271,"confidence_interval":0.936,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6783,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":894,"p95_latency_ms":1092,"p99_latency_ms":1172,"time_to_first_token_ms":195,"total_runtime_ms":993,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0271,"confidence_interval":0.9422,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1528,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":878,"p95_latency_ms":1073,"p99_latency_ms":1151,"time_to_first_token_ms":194,"total_runtime_ms":975,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0267,"confidence_interval":0.9404,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5726,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":889,"p95_latency_ms":1087,"p99_latency_ms":1166,"time_to_first_token_ms":207,"total_runtime_ms":988,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0194,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6289,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":888,"p95_latency_ms":1086,"p99_latency_ms":1165,"time_to_first_token_ms":193,"total_runtime_ms":987,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0113,"confidence_interval":0.9456,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9221,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":900,"p95_latency_ms":1100,"p99_latency_ms":1180,"time_to_first_token_ms":213,"total_runtime_ms":1000,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.9439,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6596,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":838,"p95_latency_ms":1025,"p99_latency_ms":1099,"time_to_first_token_ms":204,"total_runtime_ms":932,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0298,"confidence_interval":0.948,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6434,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":886,"p95_latency_ms":1082,"p99_latency_ms":1161,"time_to_first_token_ms":190,"total_runtime_ms":984,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0283,"confidence_interval":0.9344,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1395,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":839,"p95_latency_ms":1025,"p99_latency_ms":1100,"time_to_first_token_ms":223,"total_runtime_ms":932,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0241,"confidence_interval":0.9461,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.333,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":149,"total_runtime_ms":510,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0255,"confidence_interval":0.9342,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1403,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":531,"p95_latency_ms":649,"p99_latency_ms":697,"time_to_first_token_ms":169,"total_runtime_ms":590,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0123,"confidence_interval":0.9445,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0068,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":555,"p95_latency_ms":679,"p99_latency_ms":728,"time_to_first_token_ms":161,"total_runtime_ms":617,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0102,"confidence_interval":0.9429,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.131,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":504,"p95_latency_ms":616,"p99_latency_ms":661,"time_to_first_token_ms":150,"total_runtime_ms":560,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0271,"confidence_interval":0.9488,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9626,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":178,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0202,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7842,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":463,"p95_latency_ms":566,"p99_latency_ms":607,"time_to_first_token_ms":176,"total_runtime_ms":514,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0112,"confidence_interval":0.9428,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7396,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":514,"p95_latency_ms":628,"p99_latency_ms":674,"time_to_first_token_ms":167,"total_runtime_ms":571,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0239,"confidence_interval":0.9401,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6938,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":546,"p95_latency_ms":668,"p99_latency_ms":716,"time_to_first_token_ms":170,"total_runtime_ms":607,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0195,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2507,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":513,"p95_latency_ms":626,"p99_latency_ms":672,"time_to_first_token_ms":165,"total_runtime_ms":570,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0242,"confidence_interval":0.9455,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0037,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":471,"p95_latency_ms":576,"p99_latency_ms":618,"time_to_first_token_ms":162,"total_runtime_ms":524,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0125,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2862,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1675,"p95_latency_ms":2047,"p99_latency_ms":2196,"time_to_first_token_ms":323,"total_runtime_ms":1861,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0174,"confidence_interval":0.9315,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3087,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1692,"p95_latency_ms":2068,"p99_latency_ms":2219,"time_to_first_token_ms":343,"total_runtime_ms":1880,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0152,"confidence_interval":0.9318,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3059,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1683,"p95_latency_ms":2057,"p99_latency_ms":2206,"time_to_first_token_ms":351,"total_runtime_ms":1870,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0133,"confidence_interval":0.9417,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7174,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1718,"p95_latency_ms":2100,"p99_latency_ms":2253,"time_to_first_token_ms":345,"total_runtime_ms":1909,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.016,"confidence_interval":0.9302,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.287,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1715,"p95_latency_ms":2096,"p99_latency_ms":2249,"time_to_first_token_ms":342,"total_runtime_ms":1906,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0158,"confidence_interval":0.9334,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7679,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1775,"p95_latency_ms":2169,"p99_latency_ms":2327,"time_to_first_token_ms":325,"total_runtime_ms":1972,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0267,"confidence_interval":0.946,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1629,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1731,"p95_latency_ms":2115,"p99_latency_ms":2269,"time_to_first_token_ms":324,"total_runtime_ms":1923,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0173,"confidence_interval":0.9492,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8989,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1704,"p95_latency_ms":2082,"p99_latency_ms":2234,"time_to_first_token_ms":346,"total_runtime_ms":1893,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0164,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.673,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1710,"p95_latency_ms":2090,"p99_latency_ms":2243,"time_to_first_token_ms":329,"total_runtime_ms":1900,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0233,"confidence_interval":0.9471,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2197,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1698,"p95_latency_ms":2075,"p99_latency_ms":2226,"time_to_first_token_ms":333,"total_runtime_ms":1886,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0207,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8443,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":948,"p95_latency_ms":1159,"p99_latency_ms":1243,"time_to_first_token_ms":300,"total_runtime_ms":1053,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0294,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9975,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":971,"p95_latency_ms":1187,"p99_latency_ms":1274,"time_to_first_token_ms":297,"total_runtime_ms":1079,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0267,"confidence_interval":0.9492,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1752,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":996,"p95_latency_ms":1218,"p99_latency_ms":1306,"time_to_first_token_ms":280,"total_runtime_ms":1107,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.024,"confidence_interval":0.9444,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3615,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1010,"p95_latency_ms":1235,"p99_latency_ms":1325,"time_to_first_token_ms":291,"total_runtime_ms":1122,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0154,"confidence_interval":0.9362,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7208,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":918,"p95_latency_ms":1122,"p99_latency_ms":1203,"time_to_first_token_ms":275,"total_runtime_ms":1020,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0238,"confidence_interval":0.9475,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7701,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":999,"p95_latency_ms":1221,"p99_latency_ms":1310,"time_to_first_token_ms":301,"total_runtime_ms":1110,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0281,"confidence_interval":0.9308,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2676,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":966,"p95_latency_ms":1181,"p99_latency_ms":1267,"time_to_first_token_ms":294,"total_runtime_ms":1074,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0114,"confidence_interval":0.9465,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0319,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":970,"p95_latency_ms":1186,"p99_latency_ms":1272,"time_to_first_token_ms":295,"total_runtime_ms":1078,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.011,"confidence_interval":0.9463,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9033,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":916,"p95_latency_ms":1119,"p99_latency_ms":1200,"time_to_first_token_ms":291,"total_runtime_ms":1017,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0255,"confidence_interval":0.9453,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8795,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":926,"p95_latency_ms":1132,"p99_latency_ms":1214,"time_to_first_token_ms":297,"total_runtime_ms":1029,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.015,"confidence_interval":0.9409,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2815,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2528,"p95_latency_ms":3090,"p99_latency_ms":3314,"time_to_first_token_ms":473,"total_runtime_ms":2809,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0186,"confidence_interval":0.9496,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9172,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2514,"p95_latency_ms":3073,"p99_latency_ms":3296,"time_to_first_token_ms":452,"total_runtime_ms":2794,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0279,"confidence_interval":0.9403,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0306,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2483,"p95_latency_ms":3035,"p99_latency_ms":3256,"time_to_first_token_ms":455,"total_runtime_ms":2759,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0196,"confidence_interval":0.9356,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6238,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2538,"p95_latency_ms":3102,"p99_latency_ms":3328,"time_to_first_token_ms":473,"total_runtime_ms":2820,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0215,"confidence_interval":0.9392,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9047,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2545,"p95_latency_ms":3111,"p99_latency_ms":3337,"time_to_first_token_ms":446,"total_runtime_ms":2828,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0233,"confidence_interval":0.9424,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3936,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2536,"p95_latency_ms":3100,"p99_latency_ms":3325,"time_to_first_token_ms":471,"total_runtime_ms":2818,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0191,"confidence_interval":0.9486,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8301,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2522,"p95_latency_ms":3082,"p99_latency_ms":3306,"time_to_first_token_ms":469,"total_runtime_ms":2802,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0252,"confidence_interval":0.9387,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8243,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2562,"p95_latency_ms":3131,"p99_latency_ms":3358,"time_to_first_token_ms":452,"total_runtime_ms":2846,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0233,"confidence_interval":0.9372,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1047,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2501,"p95_latency_ms":3057,"p99_latency_ms":3279,"time_to_first_token_ms":448,"total_runtime_ms":2779,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.024,"confidence_interval":0.9457,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5053,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2571,"p95_latency_ms":3142,"p99_latency_ms":3371,"time_to_first_token_ms":447,"total_runtime_ms":2857,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0127,"confidence_interval":0.9485,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1841,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1392,"p95_latency_ms":1701,"p99_latency_ms":1825,"time_to_first_token_ms":417,"total_runtime_ms":1546,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0159,"confidence_interval":0.9468,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5225,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1371,"p95_latency_ms":1676,"p99_latency_ms":1798,"time_to_first_token_ms":420,"total_runtime_ms":1524,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0194,"confidence_interval":0.9488,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9258,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1400,"p95_latency_ms":1711,"p99_latency_ms":1835,"time_to_first_token_ms":412,"total_runtime_ms":1555,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0272,"confidence_interval":0.9414,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1981,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1448,"p95_latency_ms":1769,"p99_latency_ms":1898,"time_to_first_token_ms":423,"total_runtime_ms":1608,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0117,"confidence_interval":0.9499,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9961,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1359,"p95_latency_ms":1661,"p99_latency_ms":1781,"time_to_first_token_ms":418,"total_runtime_ms":1510,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0258,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2217,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1412,"p95_latency_ms":1726,"p99_latency_ms":1851,"time_to_first_token_ms":401,"total_runtime_ms":1569,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0147,"confidence_interval":0.9382,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4572,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1390,"p95_latency_ms":1699,"p99_latency_ms":1823,"time_to_first_token_ms":407,"total_runtime_ms":1545,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0132,"confidence_interval":0.9364,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1212,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1405,"p95_latency_ms":1717,"p99_latency_ms":1841,"time_to_first_token_ms":411,"total_runtime_ms":1561,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0224,"confidence_interval":0.9423,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3836,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1365,"p95_latency_ms":1668,"p99_latency_ms":1790,"time_to_first_token_ms":426,"total_runtime_ms":1517,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0143,"confidence_interval":0.9461,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4097,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1371,"p95_latency_ms":1675,"p99_latency_ms":1797,"time_to_first_token_ms":419,"total_runtime_ms":1523,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0257,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4129,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":530,"p95_latency_ms":647,"p99_latency_ms":694,"time_to_first_token_ms":160,"total_runtime_ms":588,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0247,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3331,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":556,"p95_latency_ms":680,"p99_latency_ms":729,"time_to_first_token_ms":158,"total_runtime_ms":618,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.02,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7366,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":495,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":152,"total_runtime_ms":549,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0244,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1066,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":547,"p95_latency_ms":668,"p99_latency_ms":717,"time_to_first_token_ms":158,"total_runtime_ms":608,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.014,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0065,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":472,"p95_latency_ms":577,"p99_latency_ms":619,"time_to_first_token_ms":164,"total_runtime_ms":525,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0131,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0738,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":511,"p95_latency_ms":624,"p99_latency_ms":670,"time_to_first_token_ms":167,"total_runtime_ms":568,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0176,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6449,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":468,"p95_latency_ms":572,"p99_latency_ms":614,"time_to_first_token_ms":163,"total_runtime_ms":520,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0257,"confidence_interval":0.9454,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4378,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":564,"p95_latency_ms":689,"p99_latency_ms":740,"time_to_first_token_ms":147,"total_runtime_ms":627,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0178,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7842,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":516,"p95_latency_ms":630,"p99_latency_ms":676,"time_to_first_token_ms":133,"total_runtime_ms":573,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0219,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0633,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":569,"p95_latency_ms":695,"p99_latency_ms":746,"time_to_first_token_ms":160,"total_runtime_ms":632,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0197,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5147,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":323,"p95_latency_ms":395,"p99_latency_ms":423,"time_to_first_token_ms":106,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0208,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6936,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":321,"p95_latency_ms":392,"p99_latency_ms":420,"time_to_first_token_ms":103,"total_runtime_ms":356,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5186,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":299,"p95_latency_ms":366,"p99_latency_ms":393,"time_to_first_token_ms":119,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.022,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9254,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":299,"p95_latency_ms":365,"p99_latency_ms":391,"time_to_first_token_ms":91,"total_runtime_ms":332,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0189,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1968,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":313,"p95_latency_ms":383,"p99_latency_ms":410,"time_to_first_token_ms":115,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0248,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5687,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":301,"p95_latency_ms":367,"p99_latency_ms":394,"time_to_first_token_ms":119,"total_runtime_ms":334,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0136,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.855,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":296,"p95_latency_ms":362,"p99_latency_ms":388,"time_to_first_token_ms":95,"total_runtime_ms":329,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0283,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9386,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":304,"p95_latency_ms":371,"p99_latency_ms":398,"time_to_first_token_ms":113,"total_runtime_ms":337,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0177,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0964,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":277,"p95_latency_ms":339,"p99_latency_ms":364,"time_to_first_token_ms":107,"total_runtime_ms":308,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0155,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5548,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":263,"p95_latency_ms":322,"p99_latency_ms":345,"time_to_first_token_ms":92,"total_runtime_ms":292,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0224,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2325,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":672,"p95_latency_ms":821,"p99_latency_ms":881,"time_to_first_token_ms":158,"total_runtime_ms":747,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0138,"confidence_interval":0.9331,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6116,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":716,"p95_latency_ms":875,"p99_latency_ms":938,"time_to_first_token_ms":175,"total_runtime_ms":795,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0113,"confidence_interval":0.9337,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0855,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":687,"p95_latency_ms":840,"p99_latency_ms":901,"time_to_first_token_ms":162,"total_runtime_ms":764,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0204,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9842,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":634,"p95_latency_ms":775,"p99_latency_ms":831,"time_to_first_token_ms":172,"total_runtime_ms":705,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0265,"confidence_interval":0.9342,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.496,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":932,"time_to_first_token_ms":173,"total_runtime_ms":789,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0142,"confidence_interval":0.9404,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8774,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":648,"p95_latency_ms":791,"p99_latency_ms":849,"time_to_first_token_ms":171,"total_runtime_ms":719,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0114,"confidence_interval":0.9316,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7015,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":673,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":180,"total_runtime_ms":747,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0151,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8597,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":645,"p95_latency_ms":788,"p99_latency_ms":845,"time_to_first_token_ms":163,"total_runtime_ms":716,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0152,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6147,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":655,"p95_latency_ms":800,"p99_latency_ms":858,"time_to_first_token_ms":174,"total_runtime_ms":727,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0249,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.046,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":727,"p95_latency_ms":888,"p99_latency_ms":953,"time_to_first_token_ms":159,"total_runtime_ms":808,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":45,"gpu_utilization_percent":32,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.018,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0686,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":351,"p95_latency_ms":428,"p99_latency_ms":460,"time_to_first_token_ms":118,"total_runtime_ms":389,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.027,"confidence_interval":0.9391,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4125,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":445,"p95_latency_ms":544,"p99_latency_ms":583,"time_to_first_token_ms":146,"total_runtime_ms":494,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0178,"confidence_interval":0.9403,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6045,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":376,"p95_latency_ms":460,"p99_latency_ms":493,"time_to_first_token_ms":123,"total_runtime_ms":418,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0184,"confidence_interval":0.9382,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9392,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":375,"p95_latency_ms":459,"p99_latency_ms":492,"time_to_first_token_ms":131,"total_runtime_ms":417,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0149,"confidence_interval":0.9436,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1839,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":409,"p95_latency_ms":500,"p99_latency_ms":536,"time_to_first_token_ms":123,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0147,"confidence_interval":0.9315,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3254,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":554,"time_to_first_token_ms":138,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0156,"confidence_interval":0.9324,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0693,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":477,"p99_latency_ms":512,"time_to_first_token_ms":126,"total_runtime_ms":434,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0102,"confidence_interval":0.9392,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.862,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":422,"p95_latency_ms":516,"p99_latency_ms":553,"time_to_first_token_ms":130,"total_runtime_ms":469,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0124,"confidence_interval":0.9409,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1787,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":578,"time_to_first_token_ms":128,"total_runtime_ms":489,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0148,"confidence_interval":0.9375,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3285,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":436,"p95_latency_ms":533,"p99_latency_ms":572,"time_to_first_token_ms":115,"total_runtime_ms":484,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":35,"gpu_utilization_percent":32,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0146,"confidence_interval":0.9368,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6528,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1259,"p95_latency_ms":1539,"p99_latency_ms":1651,"time_to_first_token_ms":268,"total_runtime_ms":1399,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0133,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5051,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1350,"p95_latency_ms":1650,"p99_latency_ms":1770,"time_to_first_token_ms":272,"total_runtime_ms":1500,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0205,"confidence_interval":0.9392,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7735,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1276,"p95_latency_ms":1560,"p99_latency_ms":1673,"time_to_first_token_ms":260,"total_runtime_ms":1418,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0202,"confidence_interval":0.931,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.439,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1356,"p95_latency_ms":1658,"p99_latency_ms":1779,"time_to_first_token_ms":288,"total_runtime_ms":1507,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0288,"confidence_interval":0.9397,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8085,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1252,"p95_latency_ms":1530,"p99_latency_ms":1642,"time_to_first_token_ms":258,"total_runtime_ms":1391,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0107,"confidence_interval":0.9484,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0166,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1283,"p95_latency_ms":1568,"p99_latency_ms":1682,"time_to_first_token_ms":286,"total_runtime_ms":1426,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9474,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4475,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1273,"p95_latency_ms":1556,"p99_latency_ms":1669,"time_to_first_token_ms":266,"total_runtime_ms":1414,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0236,"confidence_interval":0.9385,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7652,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1276,"p95_latency_ms":1560,"p99_latency_ms":1673,"time_to_first_token_ms":283,"total_runtime_ms":1418,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0235,"confidence_interval":0.9473,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.305,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1348,"p95_latency_ms":1648,"p99_latency_ms":1768,"time_to_first_token_ms":261,"total_runtime_ms":1498,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.027,"confidence_interval":0.9341,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9441,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1298,"p95_latency_ms":1587,"p99_latency_ms":1702,"time_to_first_token_ms":285,"total_runtime_ms":1443,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":45,"gpu_utilization_percent":65,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0177,"confidence_interval":0.9467,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6934,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":722,"p95_latency_ms":882,"p99_latency_ms":947,"time_to_first_token_ms":210,"total_runtime_ms":802,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.029,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3907,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":738,"p95_latency_ms":902,"p99_latency_ms":967,"time_to_first_token_ms":223,"total_runtime_ms":820,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0138,"confidence_interval":0.9475,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.713,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":759,"p95_latency_ms":927,"p99_latency_ms":995,"time_to_first_token_ms":212,"total_runtime_ms":843,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0204,"confidence_interval":0.9374,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1301,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":931,"time_to_first_token_ms":217,"total_runtime_ms":789,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.021,"confidence_interval":0.9356,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4469,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":740,"p95_latency_ms":904,"p99_latency_ms":970,"time_to_first_token_ms":229,"total_runtime_ms":822,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0225,"confidence_interval":0.948,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0382,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":788,"p95_latency_ms":963,"p99_latency_ms":1034,"time_to_first_token_ms":221,"total_runtime_ms":876,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0187,"confidence_interval":0.9418,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1202,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":712,"p95_latency_ms":870,"p99_latency_ms":934,"time_to_first_token_ms":239,"total_runtime_ms":791,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6321,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":749,"p95_latency_ms":915,"p99_latency_ms":982,"time_to_first_token_ms":235,"total_runtime_ms":832,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0174,"confidence_interval":0.9461,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7005,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":762,"p95_latency_ms":932,"p99_latency_ms":1000,"time_to_first_token_ms":224,"total_runtime_ms":847,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.011,"confidence_interval":0.9358,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5711,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":731,"p95_latency_ms":893,"p99_latency_ms":958,"time_to_first_token_ms":229,"total_runtime_ms":812,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":35,"gpu_utilization_percent":65,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0245,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6745,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1901,"p95_latency_ms":2324,"p99_latency_ms":2493,"time_to_first_token_ms":362,"total_runtime_ms":2113,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0267,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1882,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1949,"p95_latency_ms":2382,"p99_latency_ms":2555,"time_to_first_token_ms":362,"total_runtime_ms":2166,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0246,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6804,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1910,"p95_latency_ms":2334,"p99_latency_ms":2504,"time_to_first_token_ms":353,"total_runtime_ms":2122,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0157,"confidence_interval":0.9394,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.082,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1886,"p95_latency_ms":2306,"p99_latency_ms":2473,"time_to_first_token_ms":352,"total_runtime_ms":2096,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0172,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3205,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1906,"p95_latency_ms":2329,"p99_latency_ms":2498,"time_to_first_token_ms":348,"total_runtime_ms":2117,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0154,"confidence_interval":0.9489,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1169,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1902,"p95_latency_ms":2325,"p99_latency_ms":2494,"time_to_first_token_ms":377,"total_runtime_ms":2113,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0115,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5066,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1899,"p95_latency_ms":2321,"p99_latency_ms":2490,"time_to_first_token_ms":356,"total_runtime_ms":2110,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0117,"confidence_interval":0.9342,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7492,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1928,"p95_latency_ms":2356,"p99_latency_ms":2528,"time_to_first_token_ms":349,"total_runtime_ms":2142,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0136,"confidence_interval":0.948,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.441,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1879,"p95_latency_ms":2297,"p99_latency_ms":2464,"time_to_first_token_ms":368,"total_runtime_ms":2088,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0165,"confidence_interval":0.9393,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.551,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1925,"p95_latency_ms":2353,"p99_latency_ms":2524,"time_to_first_token_ms":351,"total_runtime_ms":2139,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":45,"gpu_utilization_percent":96,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0206,"confidence_interval":0.9311,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3673,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1022,"p95_latency_ms":1249,"p99_latency_ms":1340,"time_to_first_token_ms":324,"total_runtime_ms":1135,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0147,"confidence_interval":0.9429,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3435,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1107,"p95_latency_ms":1353,"p99_latency_ms":1451,"time_to_first_token_ms":305,"total_runtime_ms":1230,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0239,"confidence_interval":0.937,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0026,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1055,"p95_latency_ms":1289,"p99_latency_ms":1383,"time_to_first_token_ms":321,"total_runtime_ms":1172,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.018,"confidence_interval":0.9372,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.943,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1050,"p95_latency_ms":1283,"p99_latency_ms":1376,"time_to_first_token_ms":303,"total_runtime_ms":1166,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0283,"confidence_interval":0.9309,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4823,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1108,"p95_latency_ms":1354,"p99_latency_ms":1452,"time_to_first_token_ms":324,"total_runtime_ms":1231,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0282,"confidence_interval":0.9332,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8375,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1344,"time_to_first_token_ms":313,"total_runtime_ms":1139,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0131,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2311,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1018,"p95_latency_ms":1244,"p99_latency_ms":1334,"time_to_first_token_ms":325,"total_runtime_ms":1131,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.055,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1118,"p95_latency_ms":1366,"p99_latency_ms":1465,"time_to_first_token_ms":336,"total_runtime_ms":1242,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0149,"confidence_interval":0.9494,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1817,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1115,"p95_latency_ms":1363,"p99_latency_ms":1462,"time_to_first_token_ms":333,"total_runtime_ms":1239,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0255,"confidence_interval":0.9358,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6227,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1068,"p95_latency_ms":1305,"p99_latency_ms":1400,"time_to_first_token_ms":326,"total_runtime_ms":1186,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":35,"gpu_utilization_percent":96,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0202,"confidence_interval":0.9479,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1876,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":438,"p95_latency_ms":536,"p99_latency_ms":575,"time_to_first_token_ms":124,"total_runtime_ms":487,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.028,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4135,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":491,"p95_latency_ms":600,"p99_latency_ms":644,"time_to_first_token_ms":139,"total_runtime_ms":546,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0281,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8991,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":142,"total_runtime_ms":475,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0134,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5368,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":400,"p95_latency_ms":489,"p99_latency_ms":524,"time_to_first_token_ms":144,"total_runtime_ms":444,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0147,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0339,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":469,"p95_latency_ms":574,"p99_latency_ms":615,"time_to_first_token_ms":125,"total_runtime_ms":521,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0277,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6194,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":401,"p95_latency_ms":490,"p99_latency_ms":525,"time_to_first_token_ms":128,"total_runtime_ms":445,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0192,"confidence_interval":0.9419,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0506,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":479,"p95_latency_ms":586,"p99_latency_ms":628,"time_to_first_token_ms":122,"total_runtime_ms":532,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0151,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.876,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":548,"time_to_first_token_ms":121,"total_runtime_ms":464,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0151,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2874,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":394,"p95_latency_ms":482,"p99_latency_ms":517,"time_to_first_token_ms":121,"total_runtime_ms":438,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.027,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7841,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":482,"p95_latency_ms":590,"p99_latency_ms":632,"time_to_first_token_ms":136,"total_runtime_ms":536,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0132,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4394,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":236,"p95_latency_ms":288,"p99_latency_ms":309,"time_to_first_token_ms":107,"total_runtime_ms":262,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0259,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.199,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":406,"time_to_first_token_ms":96,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0107,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.869,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":241,"p95_latency_ms":294,"p99_latency_ms":316,"time_to_first_token_ms":104,"total_runtime_ms":268,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0187,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6059,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":258,"p95_latency_ms":315,"p99_latency_ms":338,"time_to_first_token_ms":83,"total_runtime_ms":287,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.017,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2981,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":270,"p95_latency_ms":330,"p99_latency_ms":354,"time_to_first_token_ms":95,"total_runtime_ms":300,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0208,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1612,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":250,"p95_latency_ms":306,"p99_latency_ms":328,"time_to_first_token_ms":88,"total_runtime_ms":278,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0276,"confidence_interval":0.9376,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2885,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":233,"p95_latency_ms":285,"p99_latency_ms":306,"time_to_first_token_ms":99,"total_runtime_ms":259,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0207,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1922,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":407,"time_to_first_token_ms":92,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0281,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0383,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":273,"p95_latency_ms":334,"p99_latency_ms":358,"time_to_first_token_ms":97,"total_runtime_ms":304,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0135,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9673,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":285,"p95_latency_ms":348,"p99_latency_ms":373,"time_to_first_token_ms":107,"total_runtime_ms":316,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0104,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3216,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":604,"p95_latency_ms":738,"p99_latency_ms":792,"time_to_first_token_ms":173,"total_runtime_ms":671,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0235,"confidence_interval":0.942,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6494,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":535,"p95_latency_ms":654,"p99_latency_ms":702,"time_to_first_token_ms":150,"total_runtime_ms":595,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0138,"confidence_interval":0.9352,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2115,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":564,"p95_latency_ms":689,"p99_latency_ms":740,"time_to_first_token_ms":149,"total_runtime_ms":627,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0158,"confidence_interval":0.9397,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8969,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":529,"p95_latency_ms":647,"p99_latency_ms":694,"time_to_first_token_ms":155,"total_runtime_ms":588,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.9361,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9839,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":539,"p95_latency_ms":659,"p99_latency_ms":707,"time_to_first_token_ms":169,"total_runtime_ms":599,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0205,"confidence_interval":0.9457,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6871,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":780,"time_to_first_token_ms":166,"total_runtime_ms":661,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0244,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9333,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":551,"p95_latency_ms":673,"p99_latency_ms":722,"time_to_first_token_ms":154,"total_runtime_ms":612,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0203,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9751,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":780,"time_to_first_token_ms":162,"total_runtime_ms":661,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.011,"confidence_interval":0.9477,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8279,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":548,"p95_latency_ms":670,"p99_latency_ms":719,"time_to_first_token_ms":159,"total_runtime_ms":609,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0297,"confidence_interval":0.946,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6813,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":601,"p95_latency_ms":735,"p99_latency_ms":788,"time_to_first_token_ms":151,"total_runtime_ms":668,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.886,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":288,"p95_latency_ms":352,"p99_latency_ms":377,"time_to_first_token_ms":110,"total_runtime_ms":320,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0279,"confidence_interval":0.9444,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3361,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":471,"time_to_first_token_ms":128,"total_runtime_ms":399,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0139,"confidence_interval":0.9467,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0416,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":305,"p95_latency_ms":373,"p99_latency_ms":400,"time_to_first_token_ms":115,"total_runtime_ms":339,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0215,"confidence_interval":0.9371,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5037,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":335,"p95_latency_ms":409,"p99_latency_ms":439,"time_to_first_token_ms":109,"total_runtime_ms":372,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0142,"confidence_interval":0.9447,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0744,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":379,"p95_latency_ms":463,"p99_latency_ms":497,"time_to_first_token_ms":115,"total_runtime_ms":421,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0275,"confidence_interval":0.9498,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2277,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":353,"p95_latency_ms":432,"p99_latency_ms":463,"time_to_first_token_ms":126,"total_runtime_ms":392,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0145,"confidence_interval":0.9366,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1974,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":387,"p95_latency_ms":472,"p99_latency_ms":507,"time_to_first_token_ms":113,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0211,"confidence_interval":0.9364,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3177,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":479,"time_to_first_token_ms":96,"total_runtime_ms":406,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0161,"confidence_interval":0.9472,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4575,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":380,"p95_latency_ms":464,"p99_latency_ms":498,"time_to_first_token_ms":101,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0142,"confidence_interval":0.9403,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6716,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":300,"p95_latency_ms":366,"p99_latency_ms":393,"time_to_first_token_ms":110,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0296,"confidence_interval":0.9397,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4784,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1110,"p95_latency_ms":1357,"p99_latency_ms":1456,"time_to_first_token_ms":226,"total_runtime_ms":1234,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9301,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1004,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1133,"p95_latency_ms":1385,"p99_latency_ms":1486,"time_to_first_token_ms":234,"total_runtime_ms":1259,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0239,"confidence_interval":0.9376,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5835,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1147,"p95_latency_ms":1402,"p99_latency_ms":1504,"time_to_first_token_ms":246,"total_runtime_ms":1275,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0268,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8721,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1069,"p95_latency_ms":1306,"p99_latency_ms":1401,"time_to_first_token_ms":246,"total_runtime_ms":1188,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0162,"confidence_interval":0.9375,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5828,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1128,"p95_latency_ms":1379,"p99_latency_ms":1480,"time_to_first_token_ms":248,"total_runtime_ms":1254,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0131,"confidence_interval":0.9452,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.915,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1119,"p95_latency_ms":1368,"p99_latency_ms":1467,"time_to_first_token_ms":254,"total_runtime_ms":1243,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0251,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8262,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1144,"p95_latency_ms":1399,"p99_latency_ms":1500,"time_to_first_token_ms":224,"total_runtime_ms":1272,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0267,"confidence_interval":0.9382,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.3099,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1094,"p95_latency_ms":1337,"p99_latency_ms":1434,"time_to_first_token_ms":246,"total_runtime_ms":1215,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0284,"confidence_interval":0.9428,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1669,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1103,"p95_latency_ms":1348,"p99_latency_ms":1446,"time_to_first_token_ms":255,"total_runtime_ms":1225,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0157,"confidence_interval":0.9454,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5329,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1121,"p95_latency_ms":1370,"p99_latency_ms":1469,"time_to_first_token_ms":237,"total_runtime_ms":1245,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0247,"confidence_interval":0.9374,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6213,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":582,"p95_latency_ms":712,"p99_latency_ms":763,"time_to_first_token_ms":190,"total_runtime_ms":647,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0148,"confidence_interval":0.9311,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2286,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":659,"p95_latency_ms":805,"p99_latency_ms":864,"time_to_first_token_ms":201,"total_runtime_ms":732,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0292,"confidence_interval":0.942,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8092,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":187,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0268,"confidence_interval":0.9393,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.517,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":573,"p95_latency_ms":700,"p99_latency_ms":751,"time_to_first_token_ms":187,"total_runtime_ms":636,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0194,"confidence_interval":0.9389,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3414,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":187,"total_runtime_ms":636,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0291,"confidence_interval":0.9471,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7792,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":669,"p95_latency_ms":818,"p99_latency_ms":877,"time_to_first_token_ms":194,"total_runtime_ms":743,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0249,"confidence_interval":0.939,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7933,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":610,"p95_latency_ms":745,"p99_latency_ms":799,"time_to_first_token_ms":180,"total_runtime_ms":677,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0114,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1291,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":659,"p95_latency_ms":806,"p99_latency_ms":864,"time_to_first_token_ms":194,"total_runtime_ms":732,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0118,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2168,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":650,"p95_latency_ms":794,"p99_latency_ms":852,"time_to_first_token_ms":204,"total_runtime_ms":722,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0176,"confidence_interval":0.9397,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4576,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:7b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":676,"p95_latency_ms":827,"p99_latency_ms":887,"time_to_first_token_ms":209,"total_runtime_ms":751,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0145,"confidence_interval":0.9359,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4867,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1564,"p95_latency_ms":1911,"p99_latency_ms":2050,"time_to_first_token_ms":328,"total_runtime_ms":1737,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0259,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7769,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1582,"p95_latency_ms":1934,"p99_latency_ms":2074,"time_to_first_token_ms":315,"total_runtime_ms":1758,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0289,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.571,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1573,"p95_latency_ms":1923,"p99_latency_ms":2062,"time_to_first_token_ms":301,"total_runtime_ms":1748,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0238,"confidence_interval":0.9301,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7638,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1603,"p95_latency_ms":1960,"p99_latency_ms":2102,"time_to_first_token_ms":326,"total_runtime_ms":1782,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0181,"confidence_interval":0.9408,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3658,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1580,"p95_latency_ms":1931,"p99_latency_ms":2071,"time_to_first_token_ms":313,"total_runtime_ms":1755,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0162,"confidence_interval":0.9484,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7176,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1557,"p95_latency_ms":1903,"p99_latency_ms":2041,"time_to_first_token_ms":306,"total_runtime_ms":1730,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0182,"confidence_interval":0.948,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0356,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1600,"p95_latency_ms":1956,"p99_latency_ms":2098,"time_to_first_token_ms":305,"total_runtime_ms":1778,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0175,"confidence_interval":0.9398,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9705,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1556,"p95_latency_ms":1902,"p99_latency_ms":2040,"time_to_first_token_ms":329,"total_runtime_ms":1729,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0198,"confidence_interval":0.9351,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6531,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1567,"p95_latency_ms":1915,"p99_latency_ms":2054,"time_to_first_token_ms":300,"total_runtime_ms":1741,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0231,"confidence_interval":0.9348,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9343,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1642,"p95_latency_ms":2007,"p99_latency_ms":2153,"time_to_first_token_ms":321,"total_runtime_ms":1825,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0113,"confidence_interval":0.9331,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0136,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":850,"p95_latency_ms":1039,"p99_latency_ms":1114,"time_to_first_token_ms":256,"total_runtime_ms":944,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0113,"confidence_interval":0.9385,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8226,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":892,"p95_latency_ms":1091,"p99_latency_ms":1170,"time_to_first_token_ms":274,"total_runtime_ms":992,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0158,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1659,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":922,"p95_latency_ms":1126,"p99_latency_ms":1208,"time_to_first_token_ms":273,"total_runtime_ms":1024,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0102,"confidence_interval":0.9491,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1049,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":860,"p95_latency_ms":1051,"p99_latency_ms":1127,"time_to_first_token_ms":256,"total_runtime_ms":955,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0195,"confidence_interval":0.9368,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9085,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":945,"p95_latency_ms":1155,"p99_latency_ms":1239,"time_to_first_token_ms":272,"total_runtime_ms":1050,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0212,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7386,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":855,"p95_latency_ms":1045,"p99_latency_ms":1121,"time_to_first_token_ms":281,"total_runtime_ms":950,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0187,"confidence_interval":0.936,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7459,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":923,"p95_latency_ms":1129,"p99_latency_ms":1211,"time_to_first_token_ms":259,"total_runtime_ms":1026,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0249,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1848,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":901,"p95_latency_ms":1102,"p99_latency_ms":1182,"time_to_first_token_ms":266,"total_runtime_ms":1001,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0104,"confidence_interval":0.9422,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3302,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":853,"p95_latency_ms":1043,"p99_latency_ms":1118,"time_to_first_token_ms":276,"total_runtime_ms":948,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0148,"confidence_interval":0.9464,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1745,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"classification_multiclass|high|qwen2:14b|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":921,"p95_latency_ms":1125,"p99_latency_ms":1207,"time_to_first_token_ms":263,"total_runtime_ms":1023,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0299,"confidence_interval":0.9405,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2482,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1053,"time_to_first_token_ms":211,"total_runtime_ms":892,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1954,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":833,"p95_latency_ms":1018,"p99_latency_ms":1092,"time_to_first_token_ms":201,"total_runtime_ms":925,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0118,"confidence_interval":0.9361,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.325,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":750,"p95_latency_ms":917,"p99_latency_ms":984,"time_to_first_token_ms":194,"total_runtime_ms":834,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0133,"confidence_interval":0.9428,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7046,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1010,"time_to_first_token_ms":218,"total_runtime_ms":856,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0191,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2762,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":779,"p95_latency_ms":952,"p99_latency_ms":1022,"time_to_first_token_ms":216,"total_runtime_ms":866,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0158,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3944,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":846,"p95_latency_ms":1034,"p99_latency_ms":1110,"time_to_first_token_ms":203,"total_runtime_ms":940,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.015,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9117,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":837,"p95_latency_ms":1023,"p99_latency_ms":1097,"time_to_first_token_ms":212,"total_runtime_ms":930,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0246,"confidence_interval":0.9346,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7052,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1053,"time_to_first_token_ms":194,"total_runtime_ms":892,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9897,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":813,"p95_latency_ms":993,"p99_latency_ms":1066,"time_to_first_token_ms":214,"total_runtime_ms":903,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0114,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9305,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":833,"p95_latency_ms":1018,"p99_latency_ms":1092,"time_to_first_token_ms":212,"total_runtime_ms":926,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0122,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9071,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":376,"p95_latency_ms":460,"p99_latency_ms":493,"time_to_first_token_ms":158,"total_runtime_ms":418,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0157,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8993,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":450,"p95_latency_ms":550,"p99_latency_ms":590,"time_to_first_token_ms":152,"total_runtime_ms":500,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0231,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9311,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":495,"time_to_first_token_ms":145,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0192,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.815,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":442,"p95_latency_ms":540,"p99_latency_ms":579,"time_to_first_token_ms":144,"total_runtime_ms":491,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0121,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8325,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":457,"p95_latency_ms":558,"p99_latency_ms":599,"time_to_first_token_ms":168,"total_runtime_ms":507,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0261,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1908,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":591,"time_to_first_token_ms":153,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0257,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0968,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":496,"time_to_first_token_ms":165,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0213,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3427,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":395,"p95_latency_ms":483,"p99_latency_ms":518,"time_to_first_token_ms":143,"total_runtime_ms":439,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0202,"confidence_interval":0.9357,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2001,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":461,"p95_latency_ms":563,"p99_latency_ms":604,"time_to_first_token_ms":150,"total_runtime_ms":512,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0231,"confidence_interval":0.9477,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1095,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":393,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":147,"total_runtime_ms":437,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.021,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.396,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1099,"p95_latency_ms":1343,"p99_latency_ms":1441,"time_to_first_token_ms":256,"total_runtime_ms":1221,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0245,"confidence_interval":0.9445,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9683,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1059,"p95_latency_ms":1295,"p99_latency_ms":1389,"time_to_first_token_ms":241,"total_runtime_ms":1177,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0112,"confidence_interval":0.9322,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0481,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1084,"p95_latency_ms":1325,"p99_latency_ms":1421,"time_to_first_token_ms":235,"total_runtime_ms":1205,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0292,"confidence_interval":0.9464,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3704,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1101,"p95_latency_ms":1346,"p99_latency_ms":1444,"time_to_first_token_ms":255,"total_runtime_ms":1223,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0101,"confidence_interval":0.931,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4672,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1032,"p95_latency_ms":1262,"p99_latency_ms":1354,"time_to_first_token_ms":257,"total_runtime_ms":1147,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0149,"confidence_interval":0.9392,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0604,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1102,"p95_latency_ms":1347,"p99_latency_ms":1444,"time_to_first_token_ms":240,"total_runtime_ms":1224,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0215,"confidence_interval":0.9408,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7521,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1044,"p95_latency_ms":1276,"p99_latency_ms":1369,"time_to_first_token_ms":235,"total_runtime_ms":1160,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0278,"confidence_interval":0.9461,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7047,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1095,"p95_latency_ms":1339,"p99_latency_ms":1436,"time_to_first_token_ms":262,"total_runtime_ms":1217,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0197,"confidence_interval":0.9406,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.634,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1018,"p95_latency_ms":1244,"p99_latency_ms":1334,"time_to_first_token_ms":240,"total_runtime_ms":1131,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0232,"confidence_interval":0.931,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5767,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1102,"p95_latency_ms":1347,"p99_latency_ms":1445,"time_to_first_token_ms":243,"total_runtime_ms":1225,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.019,"confidence_interval":0.9423,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3364,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":525,"p95_latency_ms":642,"p99_latency_ms":688,"time_to_first_token_ms":221,"total_runtime_ms":583,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0188,"confidence_interval":0.9395,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3834,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":608,"p95_latency_ms":743,"p99_latency_ms":797,"time_to_first_token_ms":206,"total_runtime_ms":676,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0141,"confidence_interval":0.9346,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.648,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":606,"p95_latency_ms":741,"p99_latency_ms":795,"time_to_first_token_ms":199,"total_runtime_ms":673,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0202,"confidence_interval":0.9479,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.802,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":602,"p95_latency_ms":736,"p99_latency_ms":790,"time_to_first_token_ms":198,"total_runtime_ms":669,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0212,"confidence_interval":0.9394,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2162,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":573,"p95_latency_ms":700,"p99_latency_ms":751,"time_to_first_token_ms":207,"total_runtime_ms":637,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0242,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9822,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":518,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":208,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0227,"confidence_interval":0.9398,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6241,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":607,"p95_latency_ms":741,"p99_latency_ms":795,"time_to_first_token_ms":192,"total_runtime_ms":674,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.024,"confidence_interval":0.9439,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7063,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":593,"p95_latency_ms":725,"p99_latency_ms":778,"time_to_first_token_ms":201,"total_runtime_ms":659,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0264,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.298,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":521,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":188,"total_runtime_ms":578,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0209,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1788,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":587,"p95_latency_ms":717,"p99_latency_ms":769,"time_to_first_token_ms":195,"total_runtime_ms":652,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0241,"confidence_interval":0.9392,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7693,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2036,"p95_latency_ms":2488,"p99_latency_ms":2669,"time_to_first_token_ms":440,"total_runtime_ms":2262,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0193,"confidence_interval":0.9388,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.489,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2112,"p95_latency_ms":2581,"p99_latency_ms":2768,"time_to_first_token_ms":421,"total_runtime_ms":2346,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0118,"confidence_interval":0.9357,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7404,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2032,"p95_latency_ms":2484,"p99_latency_ms":2664,"time_to_first_token_ms":423,"total_runtime_ms":2258,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0295,"confidence_interval":0.9462,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2263,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2079,"p95_latency_ms":2541,"p99_latency_ms":2726,"time_to_first_token_ms":408,"total_runtime_ms":2310,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0104,"confidence_interval":0.9344,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7976,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2025,"p95_latency_ms":2475,"p99_latency_ms":2655,"time_to_first_token_ms":415,"total_runtime_ms":2250,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0267,"confidence_interval":0.9398,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8287,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2065,"p95_latency_ms":2523,"p99_latency_ms":2707,"time_to_first_token_ms":434,"total_runtime_ms":2294,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0273,"confidence_interval":0.9359,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4527,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2128,"p95_latency_ms":2601,"p99_latency_ms":2790,"time_to_first_token_ms":431,"total_runtime_ms":2364,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0161,"confidence_interval":0.9312,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3476,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2049,"p95_latency_ms":2504,"p99_latency_ms":2686,"time_to_first_token_ms":422,"total_runtime_ms":2276,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0158,"confidence_interval":0.9353,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.246,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2082,"p95_latency_ms":2545,"p99_latency_ms":2730,"time_to_first_token_ms":418,"total_runtime_ms":2314,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9449,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0054,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2121,"p95_latency_ms":2593,"p99_latency_ms":2781,"time_to_first_token_ms":421,"total_runtime_ms":2357,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0238,"confidence_interval":0.9381,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9186,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1023,"p95_latency_ms":1250,"p99_latency_ms":1341,"time_to_first_token_ms":368,"total_runtime_ms":1136,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9453,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7005,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1055,"p95_latency_ms":1290,"p99_latency_ms":1384,"time_to_first_token_ms":395,"total_runtime_ms":1173,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0288,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5553,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1075,"p95_latency_ms":1313,"p99_latency_ms":1409,"time_to_first_token_ms":393,"total_runtime_ms":1194,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0195,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2128,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1035,"p95_latency_ms":1265,"p99_latency_ms":1357,"time_to_first_token_ms":362,"total_runtime_ms":1150,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0152,"confidence_interval":0.9451,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2567,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1043,"p95_latency_ms":1274,"p99_latency_ms":1367,"time_to_first_token_ms":392,"total_runtime_ms":1158,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9327,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2728,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1039,"p95_latency_ms":1270,"p99_latency_ms":1362,"time_to_first_token_ms":390,"total_runtime_ms":1154,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0263,"confidence_interval":0.9323,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2279,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1048,"p95_latency_ms":1281,"p99_latency_ms":1374,"time_to_first_token_ms":361,"total_runtime_ms":1164,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.936,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7932,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1104,"p95_latency_ms":1350,"p99_latency_ms":1448,"time_to_first_token_ms":381,"total_runtime_ms":1227,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0159,"confidence_interval":0.9414,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4963,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1081,"p95_latency_ms":1321,"p99_latency_ms":1417,"time_to_first_token_ms":383,"total_runtime_ms":1201,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0249,"confidence_interval":0.9458,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0721,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1056,"p95_latency_ms":1291,"p99_latency_ms":1385,"time_to_first_token_ms":388,"total_runtime_ms":1174,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0191,"confidence_interval":0.9408,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1735,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3000,"p95_latency_ms":3667,"p99_latency_ms":3933,"time_to_first_token_ms":573,"total_runtime_ms":3333,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0182,"confidence_interval":0.9436,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7978,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3008,"p95_latency_ms":3676,"p99_latency_ms":3943,"time_to_first_token_ms":605,"total_runtime_ms":3342,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.01,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6373,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3024,"p95_latency_ms":3696,"p99_latency_ms":3965,"time_to_first_token_ms":593,"total_runtime_ms":3360,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0198,"confidence_interval":0.9438,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8623,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3014,"p95_latency_ms":3684,"p99_latency_ms":3952,"time_to_first_token_ms":590,"total_runtime_ms":3349,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0285,"confidence_interval":0.9486,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.315,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3004,"p95_latency_ms":3672,"p99_latency_ms":3939,"time_to_first_token_ms":585,"total_runtime_ms":3338,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0148,"confidence_interval":0.9448,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5586,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3001,"p95_latency_ms":3668,"p99_latency_ms":3935,"time_to_first_token_ms":603,"total_runtime_ms":3335,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0235,"confidence_interval":0.9309,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5206,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3039,"p95_latency_ms":3715,"p99_latency_ms":3985,"time_to_first_token_ms":589,"total_runtime_ms":3377,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0156,"confidence_interval":0.9454,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3061,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3061,"p95_latency_ms":3742,"p99_latency_ms":4014,"time_to_first_token_ms":587,"total_runtime_ms":3401,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0174,"confidence_interval":0.9427,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7799,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3064,"p95_latency_ms":3745,"p99_latency_ms":4017,"time_to_first_token_ms":576,"total_runtime_ms":3405,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0245,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5454,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3020,"p95_latency_ms":3691,"p99_latency_ms":3960,"time_to_first_token_ms":602,"total_runtime_ms":3356,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0116,"confidence_interval":0.9431,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1616,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1558,"p95_latency_ms":1905,"p99_latency_ms":2043,"time_to_first_token_ms":549,"total_runtime_ms":1732,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.017,"confidence_interval":0.9335,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1383,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1533,"p95_latency_ms":1874,"p99_latency_ms":2010,"time_to_first_token_ms":549,"total_runtime_ms":1703,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0265,"confidence_interval":0.9309,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3145,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1521,"p95_latency_ms":1859,"p99_latency_ms":1994,"time_to_first_token_ms":529,"total_runtime_ms":1690,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0251,"confidence_interval":0.9447,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4507,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1550,"p95_latency_ms":1894,"p99_latency_ms":2032,"time_to_first_token_ms":555,"total_runtime_ms":1722,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0275,"confidence_interval":0.9399,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6605,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1599,"p95_latency_ms":1955,"p99_latency_ms":2097,"time_to_first_token_ms":528,"total_runtime_ms":1777,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0209,"confidence_interval":0.9479,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4356,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1501,"p95_latency_ms":1835,"p99_latency_ms":1968,"time_to_first_token_ms":537,"total_runtime_ms":1668,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0247,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7644,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1581,"p95_latency_ms":1933,"p99_latency_ms":2073,"time_to_first_token_ms":529,"total_runtime_ms":1757,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0173,"confidence_interval":0.9425,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5508,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1548,"p95_latency_ms":1892,"p99_latency_ms":2030,"time_to_first_token_ms":548,"total_runtime_ms":1720,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0252,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9105,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1501,"p95_latency_ms":1835,"p99_latency_ms":1968,"time_to_first_token_ms":555,"total_runtime_ms":1668,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0157,"confidence_interval":0.9487,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0047,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|low|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1568,"p95_latency_ms":1917,"p99_latency_ms":2056,"time_to_first_token_ms":531,"total_runtime_ms":1743,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0146,"confidence_interval":0.9381,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2616,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":616,"p95_latency_ms":753,"p99_latency_ms":808,"time_to_first_token_ms":171,"total_runtime_ms":685,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0118,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6638,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":563,"p95_latency_ms":688,"p99_latency_ms":738,"time_to_first_token_ms":189,"total_runtime_ms":626,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0182,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4368,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":623,"p95_latency_ms":762,"p99_latency_ms":817,"time_to_first_token_ms":189,"total_runtime_ms":692,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0167,"confidence_interval":0.9428,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7183,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":597,"p95_latency_ms":729,"p99_latency_ms":782,"time_to_first_token_ms":183,"total_runtime_ms":663,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0182,"confidence_interval":0.9423,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8819,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":618,"p95_latency_ms":756,"p99_latency_ms":811,"time_to_first_token_ms":181,"total_runtime_ms":687,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0209,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4612,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":656,"p95_latency_ms":801,"p99_latency_ms":860,"time_to_first_token_ms":160,"total_runtime_ms":728,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0267,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6524,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":802,"time_to_first_token_ms":180,"total_runtime_ms":680,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0293,"confidence_interval":0.9376,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9054,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":600,"p95_latency_ms":733,"p99_latency_ms":786,"time_to_first_token_ms":175,"total_runtime_ms":666,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0246,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5831,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":594,"p95_latency_ms":726,"p99_latency_ms":778,"time_to_first_token_ms":169,"total_runtime_ms":660,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0238,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3878,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":634,"p95_latency_ms":774,"p99_latency_ms":831,"time_to_first_token_ms":168,"total_runtime_ms":704,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.012,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7704,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":315,"p95_latency_ms":385,"p99_latency_ms":413,"time_to_first_token_ms":132,"total_runtime_ms":350,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.9455,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9503,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":348,"p95_latency_ms":425,"p99_latency_ms":456,"time_to_first_token_ms":121,"total_runtime_ms":386,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0158,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0463,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":312,"p95_latency_ms":381,"p99_latency_ms":409,"time_to_first_token_ms":145,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0145,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9818,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":315,"p95_latency_ms":385,"p99_latency_ms":413,"time_to_first_token_ms":115,"total_runtime_ms":350,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0182,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5529,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":284,"p95_latency_ms":347,"p99_latency_ms":373,"time_to_first_token_ms":143,"total_runtime_ms":316,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0239,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0714,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":506,"time_to_first_token_ms":124,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0144,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0206,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":437,"p99_latency_ms":469,"time_to_first_token_ms":126,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0151,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.431,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":334,"p95_latency_ms":409,"p99_latency_ms":438,"time_to_first_token_ms":144,"total_runtime_ms":371,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0138,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9329,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":408,"time_to_first_token_ms":143,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0203,"confidence_interval":0.9434,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.276,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":318,"p95_latency_ms":388,"p99_latency_ms":417,"time_to_first_token_ms":135,"total_runtime_ms":353,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0104,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2796,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":837,"p95_latency_ms":1022,"p99_latency_ms":1097,"time_to_first_token_ms":192,"total_runtime_ms":930,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0151,"confidence_interval":0.9411,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6621,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1030,"p99_latency_ms":1105,"time_to_first_token_ms":208,"total_runtime_ms":937,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0116,"confidence_interval":0.9394,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6892,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":791,"p95_latency_ms":966,"p99_latency_ms":1037,"time_to_first_token_ms":220,"total_runtime_ms":879,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0141,"confidence_interval":0.9337,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2285,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":828,"p95_latency_ms":1013,"p99_latency_ms":1086,"time_to_first_token_ms":197,"total_runtime_ms":920,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0143,"confidence_interval":0.9341,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3126,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":783,"p95_latency_ms":957,"p99_latency_ms":1027,"time_to_first_token_ms":223,"total_runtime_ms":870,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0147,"confidence_interval":0.9423,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4195,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":856,"p95_latency_ms":1047,"p99_latency_ms":1123,"time_to_first_token_ms":211,"total_runtime_ms":951,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0249,"confidence_interval":0.9384,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3258,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1003,"time_to_first_token_ms":201,"total_runtime_ms":850,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0219,"confidence_interval":0.9436,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4707,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":780,"p95_latency_ms":953,"p99_latency_ms":1022,"time_to_first_token_ms":218,"total_runtime_ms":866,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0292,"confidence_interval":0.933,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5679,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":802,"p95_latency_ms":980,"p99_latency_ms":1051,"time_to_first_token_ms":191,"total_runtime_ms":891,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.015,"confidence_interval":0.9317,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6183,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":845,"p95_latency_ms":1033,"p99_latency_ms":1108,"time_to_first_token_ms":208,"total_runtime_ms":939,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0105,"confidence_interval":0.9401,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4355,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":457,"p95_latency_ms":559,"p99_latency_ms":599,"time_to_first_token_ms":154,"total_runtime_ms":508,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0126,"confidence_interval":0.9469,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6596,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":414,"p95_latency_ms":506,"p99_latency_ms":543,"time_to_first_token_ms":172,"total_runtime_ms":460,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.029,"confidence_interval":0.9409,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6347,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":549,"time_to_first_token_ms":159,"total_runtime_ms":465,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0122,"confidence_interval":0.936,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2767,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":422,"p95_latency_ms":516,"p99_latency_ms":553,"time_to_first_token_ms":155,"total_runtime_ms":469,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9336,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2408,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":429,"p95_latency_ms":525,"p99_latency_ms":563,"time_to_first_token_ms":163,"total_runtime_ms":477,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0105,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5961,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":401,"p95_latency_ms":490,"p99_latency_ms":525,"time_to_first_token_ms":177,"total_runtime_ms":445,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0227,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6786,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":448,"p95_latency_ms":548,"p99_latency_ms":587,"time_to_first_token_ms":165,"total_runtime_ms":498,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0213,"confidence_interval":0.9393,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1252,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":471,"p95_latency_ms":576,"p99_latency_ms":618,"time_to_first_token_ms":179,"total_runtime_ms":524,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0233,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":384,"p95_latency_ms":470,"p99_latency_ms":504,"time_to_first_token_ms":171,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0225,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2967,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":164,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0208,"confidence_interval":0.9458,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1076,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1614,"p95_latency_ms":1973,"p99_latency_ms":2117,"time_to_first_token_ms":327,"total_runtime_ms":1794,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.934,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6282,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1548,"p95_latency_ms":1892,"p99_latency_ms":2029,"time_to_first_token_ms":322,"total_runtime_ms":1720,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0154,"confidence_interval":0.9461,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1192,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1607,"p95_latency_ms":1964,"p99_latency_ms":2106,"time_to_first_token_ms":321,"total_runtime_ms":1785,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0227,"confidence_interval":0.9431,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7345,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1604,"p95_latency_ms":1961,"p99_latency_ms":2104,"time_to_first_token_ms":321,"total_runtime_ms":1783,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0124,"confidence_interval":0.9344,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4091,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1525,"p95_latency_ms":1864,"p99_latency_ms":2000,"time_to_first_token_ms":325,"total_runtime_ms":1695,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.021,"confidence_interval":0.9439,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5403,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1603,"p95_latency_ms":1959,"p99_latency_ms":2101,"time_to_first_token_ms":334,"total_runtime_ms":1781,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0295,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5761,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1565,"p95_latency_ms":1912,"p99_latency_ms":2051,"time_to_first_token_ms":330,"total_runtime_ms":1739,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0151,"confidence_interval":0.9458,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3239,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1612,"p95_latency_ms":1970,"p99_latency_ms":2113,"time_to_first_token_ms":338,"total_runtime_ms":1791,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0102,"confidence_interval":0.9418,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5976,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1529,"p95_latency_ms":1869,"p99_latency_ms":2004,"time_to_first_token_ms":326,"total_runtime_ms":1699,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0259,"confidence_interval":0.9472,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2848,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1562,"p95_latency_ms":1909,"p99_latency_ms":2048,"time_to_first_token_ms":344,"total_runtime_ms":1736,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0142,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9076,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":859,"p95_latency_ms":1049,"p99_latency_ms":1126,"time_to_first_token_ms":284,"total_runtime_ms":954,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0248,"confidence_interval":0.9378,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2632,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":798,"p95_latency_ms":975,"p99_latency_ms":1046,"time_to_first_token_ms":277,"total_runtime_ms":887,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0133,"confidence_interval":0.9302,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2694,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":858,"p95_latency_ms":1048,"p99_latency_ms":1124,"time_to_first_token_ms":279,"total_runtime_ms":953,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0279,"confidence_interval":0.9392,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1631,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":281,"total_runtime_ms":849,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0102,"confidence_interval":0.9334,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8823,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":856,"p95_latency_ms":1047,"p99_latency_ms":1123,"time_to_first_token_ms":289,"total_runtime_ms":951,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0206,"confidence_interval":0.95,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3729,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":781,"p95_latency_ms":954,"p99_latency_ms":1024,"time_to_first_token_ms":299,"total_runtime_ms":868,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0195,"confidence_interval":0.9458,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.328,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":845,"p95_latency_ms":1033,"p99_latency_ms":1108,"time_to_first_token_ms":307,"total_runtime_ms":939,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0252,"confidence_interval":0.9365,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.261,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":786,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":300,"total_runtime_ms":873,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.029,"confidence_interval":0.942,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.873,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":787,"p95_latency_ms":962,"p99_latency_ms":1032,"time_to_first_token_ms":305,"total_runtime_ms":874,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0104,"confidence_interval":0.9464,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7063,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":766,"p95_latency_ms":936,"p99_latency_ms":1004,"time_to_first_token_ms":304,"total_runtime_ms":851,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.023,"confidence_interval":0.9427,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3921,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2251,"p95_latency_ms":2751,"p99_latency_ms":2951,"time_to_first_token_ms":464,"total_runtime_ms":2501,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0117,"confidence_interval":0.9488,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1633,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2325,"p95_latency_ms":2842,"p99_latency_ms":3048,"time_to_first_token_ms":464,"total_runtime_ms":2583,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0186,"confidence_interval":0.9463,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2921,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2301,"p95_latency_ms":2812,"p99_latency_ms":3017,"time_to_first_token_ms":464,"total_runtime_ms":2557,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0187,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9542,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2305,"p95_latency_ms":2817,"p99_latency_ms":3022,"time_to_first_token_ms":447,"total_runtime_ms":2561,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0172,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0542,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2303,"p95_latency_ms":2814,"p99_latency_ms":3019,"time_to_first_token_ms":449,"total_runtime_ms":2559,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0207,"confidence_interval":0.936,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0276,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2293,"p95_latency_ms":2803,"p99_latency_ms":3007,"time_to_first_token_ms":471,"total_runtime_ms":2548,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0221,"confidence_interval":0.9358,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1104,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2350,"p95_latency_ms":2872,"p99_latency_ms":3081,"time_to_first_token_ms":471,"total_runtime_ms":2611,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0234,"confidence_interval":0.9387,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0934,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2351,"p95_latency_ms":2873,"p99_latency_ms":3082,"time_to_first_token_ms":448,"total_runtime_ms":2612,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0213,"confidence_interval":0.9307,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3327,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2264,"p95_latency_ms":2767,"p99_latency_ms":2968,"time_to_first_token_ms":475,"total_runtime_ms":2516,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0277,"confidence_interval":0.9407,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8819,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2292,"p95_latency_ms":2801,"p99_latency_ms":3005,"time_to_first_token_ms":471,"total_runtime_ms":2546,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0143,"confidence_interval":0.9337,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2318,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1139,"p95_latency_ms":1393,"p99_latency_ms":1494,"time_to_first_token_ms":426,"total_runtime_ms":1266,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.016,"confidence_interval":0.9401,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3624,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1141,"p95_latency_ms":1394,"p99_latency_ms":1495,"time_to_first_token_ms":432,"total_runtime_ms":1267,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0223,"confidence_interval":0.9377,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5514,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1160,"p95_latency_ms":1418,"p99_latency_ms":1521,"time_to_first_token_ms":432,"total_runtime_ms":1289,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0155,"confidence_interval":0.9321,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7072,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1179,"p95_latency_ms":1441,"p99_latency_ms":1546,"time_to_first_token_ms":403,"total_runtime_ms":1310,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0161,"confidence_interval":0.9423,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1375,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1126,"p95_latency_ms":1377,"p99_latency_ms":1477,"time_to_first_token_ms":406,"total_runtime_ms":1252,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0246,"confidence_interval":0.936,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7951,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1150,"p95_latency_ms":1406,"p99_latency_ms":1508,"time_to_first_token_ms":414,"total_runtime_ms":1278,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0243,"confidence_interval":0.9382,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8403,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1154,"p95_latency_ms":1410,"p99_latency_ms":1512,"time_to_first_token_ms":405,"total_runtime_ms":1282,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.02,"confidence_interval":0.9342,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3693,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1210,"p95_latency_ms":1479,"p99_latency_ms":1587,"time_to_first_token_ms":428,"total_runtime_ms":1345,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0108,"confidence_interval":0.9407,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4149,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1139,"p95_latency_ms":1392,"p99_latency_ms":1494,"time_to_first_token_ms":402,"total_runtime_ms":1266,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9373,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3265,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1126,"p95_latency_ms":1377,"p99_latency_ms":1477,"time_to_first_token_ms":409,"total_runtime_ms":1252,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.022,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3806,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":511,"p95_latency_ms":625,"p99_latency_ms":670,"time_to_first_token_ms":162,"total_runtime_ms":568,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0103,"confidence_interval":0.9362,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4854,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":434,"p95_latency_ms":530,"p99_latency_ms":568,"time_to_first_token_ms":136,"total_runtime_ms":482,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0272,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8915,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":486,"p95_latency_ms":594,"p99_latency_ms":638,"time_to_first_token_ms":152,"total_runtime_ms":540,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.01,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4072,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":471,"p95_latency_ms":576,"p99_latency_ms":618,"time_to_first_token_ms":156,"total_runtime_ms":523,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0233,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3488,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":528,"p95_latency_ms":645,"p99_latency_ms":692,"time_to_first_token_ms":164,"total_runtime_ms":587,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.018,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.534,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":463,"p95_latency_ms":566,"p99_latency_ms":607,"time_to_first_token_ms":155,"total_runtime_ms":514,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0209,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5809,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":441,"p95_latency_ms":539,"p99_latency_ms":578,"time_to_first_token_ms":161,"total_runtime_ms":490,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0259,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9833,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":437,"p95_latency_ms":535,"p99_latency_ms":574,"time_to_first_token_ms":154,"total_runtime_ms":486,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.02,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5718,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":496,"p95_latency_ms":607,"p99_latency_ms":651,"time_to_first_token_ms":159,"total_runtime_ms":552,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0167,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4014,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":497,"p95_latency_ms":607,"p99_latency_ms":651,"time_to_first_token_ms":165,"total_runtime_ms":552,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0186,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5375,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":307,"p95_latency_ms":375,"p99_latency_ms":402,"time_to_first_token_ms":103,"total_runtime_ms":341,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6476,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":295,"p95_latency_ms":361,"p99_latency_ms":387,"time_to_first_token_ms":93,"total_runtime_ms":328,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0102,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1352,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":260,"p95_latency_ms":317,"p99_latency_ms":340,"time_to_first_token_ms":98,"total_runtime_ms":288,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0299,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0784,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":277,"p95_latency_ms":339,"p99_latency_ms":364,"time_to_first_token_ms":99,"total_runtime_ms":308,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9305,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3772,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":283,"p95_latency_ms":346,"p99_latency_ms":371,"time_to_first_token_ms":101,"total_runtime_ms":315,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0122,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4535,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":234,"p95_latency_ms":287,"p99_latency_ms":307,"time_to_first_token_ms":116,"total_runtime_ms":261,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0248,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1933,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":275,"p95_latency_ms":336,"p99_latency_ms":361,"time_to_first_token_ms":104,"total_runtime_ms":306,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0112,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1727,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":234,"p95_latency_ms":287,"p99_latency_ms":307,"time_to_first_token_ms":121,"total_runtime_ms":261,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0126,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3424,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":219,"p95_latency_ms":268,"p99_latency_ms":287,"time_to_first_token_ms":113,"total_runtime_ms":243,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0178,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1314,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":318,"p95_latency_ms":389,"p99_latency_ms":417,"time_to_first_token_ms":112,"total_runtime_ms":354,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0156,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7913,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":658,"p95_latency_ms":804,"p99_latency_ms":862,"time_to_first_token_ms":189,"total_runtime_ms":731,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.948,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7763,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":623,"p95_latency_ms":761,"p99_latency_ms":816,"time_to_first_token_ms":163,"total_runtime_ms":692,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0184,"confidence_interval":0.948,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9869,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":624,"p95_latency_ms":763,"p99_latency_ms":818,"time_to_first_token_ms":167,"total_runtime_ms":694,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0297,"confidence_interval":0.9333,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.415,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":165,"total_runtime_ms":741,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0149,"confidence_interval":0.9422,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1665,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":614,"p95_latency_ms":750,"p99_latency_ms":805,"time_to_first_token_ms":190,"total_runtime_ms":682,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0159,"confidence_interval":0.9418,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9849,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":772,"time_to_first_token_ms":190,"total_runtime_ms":655,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0299,"confidence_interval":0.9345,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9337,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":843,"time_to_first_token_ms":165,"total_runtime_ms":714,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0279,"confidence_interval":0.9401,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2416,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":617,"p95_latency_ms":754,"p99_latency_ms":809,"time_to_first_token_ms":171,"total_runtime_ms":686,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0191,"confidence_interval":0.95,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8417,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":663,"p95_latency_ms":810,"p99_latency_ms":869,"time_to_first_token_ms":178,"total_runtime_ms":737,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0267,"confidence_interval":0.945,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4163,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":603,"p95_latency_ms":737,"p99_latency_ms":791,"time_to_first_token_ms":158,"total_runtime_ms":670,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0155,"confidence_interval":0.9496,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1055,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":296,"p95_latency_ms":361,"p99_latency_ms":388,"time_to_first_token_ms":143,"total_runtime_ms":328,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0125,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3486,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":424,"time_to_first_token_ms":142,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0293,"confidence_interval":0.947,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1271,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":335,"p95_latency_ms":409,"p99_latency_ms":439,"time_to_first_token_ms":132,"total_runtime_ms":372,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0247,"confidence_interval":0.9328,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7699,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":299,"p95_latency_ms":365,"p99_latency_ms":392,"time_to_first_token_ms":115,"total_runtime_ms":332,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0138,"confidence_interval":0.9397,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0647,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":330,"p95_latency_ms":403,"p99_latency_ms":432,"time_to_first_token_ms":146,"total_runtime_ms":366,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0175,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.008,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":323,"p95_latency_ms":395,"p99_latency_ms":424,"time_to_first_token_ms":124,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0294,"confidence_interval":0.95,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.102,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":302,"p95_latency_ms":369,"p99_latency_ms":396,"time_to_first_token_ms":131,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0219,"confidence_interval":0.9442,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5575,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":388,"p95_latency_ms":474,"p99_latency_ms":509,"time_to_first_token_ms":138,"total_runtime_ms":431,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0214,"confidence_interval":0.9437,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6183,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":325,"p95_latency_ms":397,"p99_latency_ms":426,"time_to_first_token_ms":120,"total_runtime_ms":361,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.945,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7927,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":362,"p95_latency_ms":443,"p99_latency_ms":475,"time_to_first_token_ms":131,"total_runtime_ms":402,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0173,"confidence_interval":0.9324,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.008,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1189,"p95_latency_ms":1453,"p99_latency_ms":1559,"time_to_first_token_ms":275,"total_runtime_ms":1321,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0129,"confidence_interval":0.9427,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0458,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1194,"p95_latency_ms":1460,"p99_latency_ms":1566,"time_to_first_token_ms":277,"total_runtime_ms":1327,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0186,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7035,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1162,"p95_latency_ms":1420,"p99_latency_ms":1523,"time_to_first_token_ms":283,"total_runtime_ms":1291,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.013,"confidence_interval":0.9412,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2351,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1176,"p95_latency_ms":1437,"p99_latency_ms":1542,"time_to_first_token_ms":266,"total_runtime_ms":1306,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.935,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0803,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1152,"p95_latency_ms":1408,"p99_latency_ms":1511,"time_to_first_token_ms":262,"total_runtime_ms":1280,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.9416,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8766,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1201,"p95_latency_ms":1468,"p99_latency_ms":1575,"time_to_first_token_ms":258,"total_runtime_ms":1334,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0289,"confidence_interval":0.9318,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.752,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1202,"p95_latency_ms":1469,"p99_latency_ms":1576,"time_to_first_token_ms":263,"total_runtime_ms":1336,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0154,"confidence_interval":0.9449,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9213,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1194,"p95_latency_ms":1459,"p99_latency_ms":1565,"time_to_first_token_ms":257,"total_runtime_ms":1326,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0187,"confidence_interval":0.9451,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3253,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1162,"p95_latency_ms":1420,"p99_latency_ms":1523,"time_to_first_token_ms":274,"total_runtime_ms":1291,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0225,"confidence_interval":0.9365,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4867,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1183,"p95_latency_ms":1446,"p99_latency_ms":1552,"time_to_first_token_ms":271,"total_runtime_ms":1315,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0281,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3691,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":616,"p95_latency_ms":753,"p99_latency_ms":808,"time_to_first_token_ms":214,"total_runtime_ms":685,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0176,"confidence_interval":0.9461,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7504,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":618,"p95_latency_ms":755,"p99_latency_ms":810,"time_to_first_token_ms":237,"total_runtime_ms":686,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0291,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9706,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":757,"time_to_first_token_ms":230,"total_runtime_ms":641,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0153,"confidence_interval":0.9305,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9551,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":659,"p95_latency_ms":805,"p99_latency_ms":863,"time_to_first_token_ms":227,"total_runtime_ms":732,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0292,"confidence_interval":0.9424,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3302,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":574,"p95_latency_ms":702,"p99_latency_ms":753,"time_to_first_token_ms":233,"total_runtime_ms":638,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.028,"confidence_interval":0.9395,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3544,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":675,"p95_latency_ms":825,"p99_latency_ms":885,"time_to_first_token_ms":238,"total_runtime_ms":750,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0266,"confidence_interval":0.9431,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9519,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":676,"p95_latency_ms":826,"p99_latency_ms":886,"time_to_first_token_ms":237,"total_runtime_ms":751,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0177,"confidence_interval":0.9346,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7524,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":612,"p95_latency_ms":749,"p99_latency_ms":803,"time_to_first_token_ms":233,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.9374,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7784,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":667,"p95_latency_ms":816,"p99_latency_ms":875,"time_to_first_token_ms":227,"total_runtime_ms":741,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0122,"confidence_interval":0.9423,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8788,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":666,"p95_latency_ms":814,"p99_latency_ms":874,"time_to_first_token_ms":221,"total_runtime_ms":740,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0135,"confidence_interval":0.9494,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0611,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1731,"p95_latency_ms":2116,"p99_latency_ms":2269,"time_to_first_token_ms":378,"total_runtime_ms":1923,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0156,"confidence_interval":0.9382,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.859,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1749,"p95_latency_ms":2138,"p99_latency_ms":2294,"time_to_first_token_ms":366,"total_runtime_ms":1944,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0294,"confidence_interval":0.9406,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4332,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1687,"p95_latency_ms":2062,"p99_latency_ms":2211,"time_to_first_token_ms":357,"total_runtime_ms":1874,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0194,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5448,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1711,"p95_latency_ms":2091,"p99_latency_ms":2243,"time_to_first_token_ms":361,"total_runtime_ms":1901,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0169,"confidence_interval":0.9491,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.839,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1772,"p95_latency_ms":2166,"p99_latency_ms":2324,"time_to_first_token_ms":361,"total_runtime_ms":1969,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0248,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5445,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1761,"p95_latency_ms":2153,"p99_latency_ms":2309,"time_to_first_token_ms":377,"total_runtime_ms":1957,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0109,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4021,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1767,"p95_latency_ms":2159,"p99_latency_ms":2316,"time_to_first_token_ms":368,"total_runtime_ms":1963,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0281,"confidence_interval":0.9316,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0157,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1700,"p95_latency_ms":2078,"p99_latency_ms":2229,"time_to_first_token_ms":360,"total_runtime_ms":1889,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0294,"confidence_interval":0.9333,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6402,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1789,"p95_latency_ms":2187,"p99_latency_ms":2346,"time_to_first_token_ms":361,"total_runtime_ms":1988,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0152,"confidence_interval":0.9408,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4297,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1754,"p95_latency_ms":2144,"p99_latency_ms":2300,"time_to_first_token_ms":361,"total_runtime_ms":1949,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0219,"confidence_interval":0.9357,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2125,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":923,"p95_latency_ms":1128,"p99_latency_ms":1210,"time_to_first_token_ms":321,"total_runtime_ms":1026,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4234,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":886,"p95_latency_ms":1083,"p99_latency_ms":1161,"time_to_first_token_ms":305,"total_runtime_ms":984,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.021,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8822,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":910,"p95_latency_ms":1113,"p99_latency_ms":1194,"time_to_first_token_ms":315,"total_runtime_ms":1011,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0132,"confidence_interval":0.9427,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4455,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":923,"p95_latency_ms":1128,"p99_latency_ms":1210,"time_to_first_token_ms":307,"total_runtime_ms":1026,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0184,"confidence_interval":0.9342,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6882,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":936,"p95_latency_ms":1144,"p99_latency_ms":1227,"time_to_first_token_ms":327,"total_runtime_ms":1040,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0142,"confidence_interval":0.9353,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7246,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":870,"p95_latency_ms":1064,"p99_latency_ms":1141,"time_to_first_token_ms":311,"total_runtime_ms":967,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0162,"confidence_interval":0.9473,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1014,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":947,"p95_latency_ms":1158,"p99_latency_ms":1242,"time_to_first_token_ms":331,"total_runtime_ms":1053,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0181,"confidence_interval":0.9443,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8177,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":916,"p95_latency_ms":1119,"p99_latency_ms":1200,"time_to_first_token_ms":331,"total_runtime_ms":1017,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0266,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7761,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":933,"p95_latency_ms":1141,"p99_latency_ms":1224,"time_to_first_token_ms":336,"total_runtime_ms":1037,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0113,"confidence_interval":0.949,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9503,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":870,"p95_latency_ms":1064,"p99_latency_ms":1141,"time_to_first_token_ms":328,"total_runtime_ms":967,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.016,"confidence_interval":0.9454,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.396,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":131,"total_runtime_ms":456,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0213,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5541,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":437,"p95_latency_ms":534,"p99_latency_ms":572,"time_to_first_token_ms":124,"total_runtime_ms":485,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0275,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8008,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":370,"p95_latency_ms":452,"p99_latency_ms":485,"time_to_first_token_ms":125,"total_runtime_ms":411,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0181,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.571,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":440,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":130,"total_runtime_ms":488,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1723,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":375,"p95_latency_ms":458,"p99_latency_ms":492,"time_to_first_token_ms":152,"total_runtime_ms":417,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5346,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":373,"p95_latency_ms":456,"p99_latency_ms":490,"time_to_first_token_ms":122,"total_runtime_ms":415,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.021,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2479,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":376,"p95_latency_ms":459,"p99_latency_ms":493,"time_to_first_token_ms":131,"total_runtime_ms":418,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0141,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6959,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":400,"p95_latency_ms":489,"p99_latency_ms":525,"time_to_first_token_ms":121,"total_runtime_ms":445,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0265,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1246,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":404,"p95_latency_ms":493,"p99_latency_ms":529,"time_to_first_token_ms":138,"total_runtime_ms":448,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8767,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":450,"p95_latency_ms":550,"p99_latency_ms":590,"time_to_first_token_ms":121,"total_runtime_ms":500,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0297,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0055,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":205,"p95_latency_ms":250,"p99_latency_ms":269,"time_to_first_token_ms":78,"total_runtime_ms":228,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3143,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":279,"p95_latency_ms":341,"p99_latency_ms":366,"time_to_first_token_ms":93,"total_runtime_ms":310,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.028,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9765,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":283,"p95_latency_ms":346,"p99_latency_ms":371,"time_to_first_token_ms":103,"total_runtime_ms":314,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0124,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1384,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":236,"p95_latency_ms":288,"p99_latency_ms":309,"time_to_first_token_ms":90,"total_runtime_ms":262,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0257,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7306,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":177,"p95_latency_ms":217,"p99_latency_ms":232,"time_to_first_token_ms":108,"total_runtime_ms":197,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.023,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4026,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":214,"p95_latency_ms":261,"p99_latency_ms":280,"time_to_first_token_ms":92,"total_runtime_ms":238,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0214,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6901,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":291,"p99_latency_ms":312,"time_to_first_token_ms":77,"total_runtime_ms":264,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0185,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3308,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":270,"p95_latency_ms":330,"p99_latency_ms":354,"time_to_first_token_ms":76,"total_runtime_ms":300,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0285,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6284,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":211,"p95_latency_ms":258,"p99_latency_ms":276,"time_to_first_token_ms":83,"total_runtime_ms":234,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0142,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1346,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":198,"p95_latency_ms":242,"p99_latency_ms":259,"time_to_first_token_ms":98,"total_runtime_ms":220,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0278,"confidence_interval":0.9455,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8236,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":160,"total_runtime_ms":575,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0213,"confidence_interval":0.9375,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7514,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":166,"total_runtime_ms":636,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0145,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3246,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":573,"p95_latency_ms":700,"p99_latency_ms":751,"time_to_first_token_ms":141,"total_runtime_ms":636,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0264,"confidence_interval":0.9433,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1399,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":474,"p95_latency_ms":579,"p99_latency_ms":621,"time_to_first_token_ms":161,"total_runtime_ms":527,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0171,"confidence_interval":0.9458,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8492,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":494,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":154,"total_runtime_ms":549,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0228,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4787,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":549,"p95_latency_ms":671,"p99_latency_ms":719,"time_to_first_token_ms":144,"total_runtime_ms":610,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0294,"confidence_interval":0.9328,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3762,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":521,"p95_latency_ms":637,"p99_latency_ms":683,"time_to_first_token_ms":156,"total_runtime_ms":579,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0103,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.865,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":526,"p95_latency_ms":642,"p99_latency_ms":689,"time_to_first_token_ms":149,"total_runtime_ms":584,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0202,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0691,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":631,"time_to_first_token_ms":152,"total_runtime_ms":535,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0102,"confidence_interval":0.9301,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9463,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":542,"p95_latency_ms":663,"p99_latency_ms":711,"time_to_first_token_ms":153,"total_runtime_ms":602,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0154,"confidence_interval":0.9347,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6887,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":278,"p95_latency_ms":339,"p99_latency_ms":364,"time_to_first_token_ms":129,"total_runtime_ms":308,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0234,"confidence_interval":0.9355,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0473,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":290,"p99_latency_ms":312,"time_to_first_token_ms":105,"total_runtime_ms":264,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0221,"confidence_interval":0.9379,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5423,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":239,"p95_latency_ms":292,"p99_latency_ms":313,"time_to_first_token_ms":106,"total_runtime_ms":266,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0117,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2852,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":264,"p95_latency_ms":323,"p99_latency_ms":347,"time_to_first_token_ms":116,"total_runtime_ms":294,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0222,"confidence_interval":0.9333,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4612,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":278,"p95_latency_ms":340,"p99_latency_ms":365,"time_to_first_token_ms":99,"total_runtime_ms":309,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.017,"confidence_interval":0.95,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6441,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":288,"p95_latency_ms":352,"p99_latency_ms":377,"time_to_first_token_ms":117,"total_runtime_ms":320,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0197,"confidence_interval":0.9453,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3753,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":299,"p95_latency_ms":366,"p99_latency_ms":393,"time_to_first_token_ms":100,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0157,"confidence_interval":0.9335,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9513,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":342,"p95_latency_ms":418,"p99_latency_ms":448,"time_to_first_token_ms":131,"total_runtime_ms":380,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0188,"confidence_interval":0.9387,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3822,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":301,"p95_latency_ms":368,"p99_latency_ms":395,"time_to_first_token_ms":112,"total_runtime_ms":335,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0268,"confidence_interval":0.9314,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4862,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":287,"p95_latency_ms":351,"p99_latency_ms":376,"time_to_first_token_ms":101,"total_runtime_ms":319,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0205,"confidence_interval":0.94,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4438,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1007,"p95_latency_ms":1231,"p99_latency_ms":1320,"time_to_first_token_ms":246,"total_runtime_ms":1119,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0198,"confidence_interval":0.9488,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8174,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1046,"p95_latency_ms":1278,"p99_latency_ms":1371,"time_to_first_token_ms":231,"total_runtime_ms":1162,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0104,"confidence_interval":0.9318,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6322,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1027,"p95_latency_ms":1255,"p99_latency_ms":1346,"time_to_first_token_ms":253,"total_runtime_ms":1141,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0146,"confidence_interval":0.9347,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1108,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1345,"time_to_first_token_ms":224,"total_runtime_ms":1139,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0224,"confidence_interval":0.9417,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6997,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1026,"p95_latency_ms":1254,"p99_latency_ms":1345,"time_to_first_token_ms":248,"total_runtime_ms":1140,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0267,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0876,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":955,"p95_latency_ms":1167,"p99_latency_ms":1252,"time_to_first_token_ms":254,"total_runtime_ms":1061,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0173,"confidence_interval":0.9357,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9221,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1017,"p95_latency_ms":1243,"p99_latency_ms":1334,"time_to_first_token_ms":246,"total_runtime_ms":1130,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0297,"confidence_interval":0.9435,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.242,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1032,"p95_latency_ms":1261,"p99_latency_ms":1353,"time_to_first_token_ms":230,"total_runtime_ms":1147,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0166,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8695,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":970,"p95_latency_ms":1186,"p99_latency_ms":1272,"time_to_first_token_ms":246,"total_runtime_ms":1078,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0271,"confidence_interval":0.9379,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.3075,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":989,"p95_latency_ms":1208,"p99_latency_ms":1296,"time_to_first_token_ms":222,"total_runtime_ms":1098,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0148,"confidence_interval":0.944,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4718,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":561,"p95_latency_ms":686,"p99_latency_ms":736,"time_to_first_token_ms":211,"total_runtime_ms":624,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0119,"confidence_interval":0.9374,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7221,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":504,"p95_latency_ms":616,"p99_latency_ms":661,"time_to_first_token_ms":193,"total_runtime_ms":560,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0202,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8238,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":212,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0111,"confidence_interval":0.9428,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8512,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":496,"p95_latency_ms":607,"p99_latency_ms":651,"time_to_first_token_ms":193,"total_runtime_ms":552,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0156,"confidence_interval":0.9473,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2057,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":581,"p99_latency_ms":623,"time_to_first_token_ms":179,"total_runtime_ms":528,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0289,"confidence_interval":0.9344,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0457,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":537,"p95_latency_ms":656,"p99_latency_ms":704,"time_to_first_token_ms":181,"total_runtime_ms":597,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0127,"confidence_interval":0.9437,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2354,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":544,"p95_latency_ms":665,"p99_latency_ms":713,"time_to_first_token_ms":201,"total_runtime_ms":604,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0174,"confidence_interval":0.949,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7244,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":538,"p95_latency_ms":657,"p99_latency_ms":705,"time_to_first_token_ms":211,"total_runtime_ms":598,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0198,"confidence_interval":0.9401,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3315,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":480,"p95_latency_ms":586,"p99_latency_ms":629,"time_to_first_token_ms":199,"total_runtime_ms":533,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0191,"confidence_interval":0.9329,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3396,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":551,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":210,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0102,"confidence_interval":0.9463,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4854,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1418,"p95_latency_ms":1733,"p99_latency_ms":1859,"time_to_first_token_ms":333,"total_runtime_ms":1575,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0122,"confidence_interval":0.9395,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8797,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1508,"p95_latency_ms":1844,"p99_latency_ms":1978,"time_to_first_token_ms":317,"total_runtime_ms":1676,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0138,"confidence_interval":0.9312,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0723,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1408,"p95_latency_ms":1721,"p99_latency_ms":1846,"time_to_first_token_ms":314,"total_runtime_ms":1564,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0161,"confidence_interval":0.9321,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8659,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1444,"p95_latency_ms":1764,"p99_latency_ms":1893,"time_to_first_token_ms":326,"total_runtime_ms":1604,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.9456,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0324,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1448,"p95_latency_ms":1770,"p99_latency_ms":1899,"time_to_first_token_ms":332,"total_runtime_ms":1609,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0166,"confidence_interval":0.9473,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5207,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1497,"p95_latency_ms":1830,"p99_latency_ms":1963,"time_to_first_token_ms":322,"total_runtime_ms":1663,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0112,"confidence_interval":0.9384,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.69,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1485,"p95_latency_ms":1816,"p99_latency_ms":1948,"time_to_first_token_ms":310,"total_runtime_ms":1651,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0139,"confidence_interval":0.938,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7694,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1500,"p95_latency_ms":1834,"p99_latency_ms":1967,"time_to_first_token_ms":308,"total_runtime_ms":1667,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0251,"confidence_interval":0.9361,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7322,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1510,"p95_latency_ms":1846,"p99_latency_ms":1980,"time_to_first_token_ms":301,"total_runtime_ms":1678,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0235,"confidence_interval":0.9488,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9219,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1458,"p95_latency_ms":1782,"p99_latency_ms":1911,"time_to_first_token_ms":308,"total_runtime_ms":1620,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0177,"confidence_interval":0.9332,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7079,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":759,"p95_latency_ms":927,"p99_latency_ms":995,"time_to_first_token_ms":269,"total_runtime_ms":843,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9372,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5342,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":763,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":262,"total_runtime_ms":848,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0156,"confidence_interval":0.9389,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8556,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":706,"p95_latency_ms":863,"p99_latency_ms":926,"time_to_first_token_ms":283,"total_runtime_ms":784,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0135,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1803,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":791,"p95_latency_ms":967,"p99_latency_ms":1037,"time_to_first_token_ms":275,"total_runtime_ms":879,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0127,"confidence_interval":0.9321,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1549,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":773,"p95_latency_ms":945,"p99_latency_ms":1014,"time_to_first_token_ms":268,"total_runtime_ms":859,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0171,"confidence_interval":0.9351,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6437,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":724,"p95_latency_ms":885,"p99_latency_ms":949,"time_to_first_token_ms":269,"total_runtime_ms":804,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0299,"confidence_interval":0.9452,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0979,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":798,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":285,"total_runtime_ms":887,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0216,"confidence_interval":0.9412,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.899,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":795,"p95_latency_ms":972,"p99_latency_ms":1042,"time_to_first_token_ms":273,"total_runtime_ms":883,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0228,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7838,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1055,"time_to_first_token_ms":259,"total_runtime_ms":894,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0189,"confidence_interval":0.945,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9358,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_entities|high|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":773,"p95_latency_ms":944,"p99_latency_ms":1013,"time_to_first_token_ms":280,"total_runtime_ms":858,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0241,"confidence_interval":0.9425,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2343,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":950,"p95_latency_ms":1162,"p99_latency_ms":1246,"time_to_first_token_ms":189,"total_runtime_ms":1056,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0118,"confidence_interval":0.9352,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3058,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":970,"p95_latency_ms":1185,"p99_latency_ms":1271,"time_to_first_token_ms":194,"total_runtime_ms":1077,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0167,"confidence_interval":0.9379,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7261,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":984,"p95_latency_ms":1202,"p99_latency_ms":1290,"time_to_first_token_ms":194,"total_runtime_ms":1093,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0147,"confidence_interval":0.9314,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7606,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":891,"p95_latency_ms":1088,"p99_latency_ms":1168,"time_to_first_token_ms":207,"total_runtime_ms":990,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0106,"confidence_interval":0.9425,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6335,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":947,"p95_latency_ms":1158,"p99_latency_ms":1242,"time_to_first_token_ms":189,"total_runtime_ms":1052,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0123,"confidence_interval":0.94,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4856,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":934,"p95_latency_ms":1141,"p99_latency_ms":1224,"time_to_first_token_ms":189,"total_runtime_ms":1038,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0225,"confidence_interval":0.9429,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6092,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":891,"p95_latency_ms":1089,"p99_latency_ms":1169,"time_to_first_token_ms":193,"total_runtime_ms":990,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.019,"confidence_interval":0.9374,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0491,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":960,"p95_latency_ms":1174,"p99_latency_ms":1259,"time_to_first_token_ms":194,"total_runtime_ms":1067,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0115,"confidence_interval":0.9371,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9719,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":965,"p95_latency_ms":1179,"p99_latency_ms":1265,"time_to_first_token_ms":195,"total_runtime_ms":1072,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0177,"confidence_interval":0.9483,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1418,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":930,"p95_latency_ms":1137,"p99_latency_ms":1220,"time_to_first_token_ms":220,"total_runtime_ms":1034,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0278,"confidence_interval":0.949,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5499,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":549,"p95_latency_ms":671,"p99_latency_ms":720,"time_to_first_token_ms":154,"total_runtime_ms":610,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0151,"confidence_interval":0.9356,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0579,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":554,"p95_latency_ms":678,"p99_latency_ms":727,"time_to_first_token_ms":156,"total_runtime_ms":616,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0269,"confidence_interval":0.9346,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3988,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":578,"p95_latency_ms":707,"p99_latency_ms":758,"time_to_first_token_ms":163,"total_runtime_ms":642,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0107,"confidence_interval":0.9375,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.237,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":569,"p95_latency_ms":695,"p99_latency_ms":745,"time_to_first_token_ms":151,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0154,"confidence_interval":0.9301,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1079,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":159,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0244,"confidence_interval":0.9403,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8771,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":785,"time_to_first_token_ms":148,"total_runtime_ms":666,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0249,"confidence_interval":0.9309,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7989,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":534,"p95_latency_ms":652,"p99_latency_ms":700,"time_to_first_token_ms":159,"total_runtime_ms":593,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0237,"confidence_interval":0.9401,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7851,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":519,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":155,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0207,"confidence_interval":0.9433,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1771,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":549,"p95_latency_ms":671,"p99_latency_ms":719,"time_to_first_token_ms":178,"total_runtime_ms":610,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0107,"confidence_interval":0.9427,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6009,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":534,"p95_latency_ms":653,"p99_latency_ms":700,"time_to_first_token_ms":153,"total_runtime_ms":593,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.026,"confidence_interval":0.939,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0808,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1253,"p95_latency_ms":1532,"p99_latency_ms":1643,"time_to_first_token_ms":261,"total_runtime_ms":1392,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.024,"confidence_interval":0.9409,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0288,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1289,"p95_latency_ms":1576,"p99_latency_ms":1691,"time_to_first_token_ms":261,"total_runtime_ms":1433,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.9437,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4422,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1246,"p95_latency_ms":1523,"p99_latency_ms":1633,"time_to_first_token_ms":236,"total_runtime_ms":1384,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.026,"confidence_interval":0.9303,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8695,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1229,"p95_latency_ms":1502,"p99_latency_ms":1611,"time_to_first_token_ms":240,"total_runtime_ms":1365,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0138,"confidence_interval":0.9446,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9571,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1219,"p95_latency_ms":1489,"p99_latency_ms":1598,"time_to_first_token_ms":254,"total_runtime_ms":1354,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0129,"confidence_interval":0.9319,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8171,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1282,"p95_latency_ms":1566,"p99_latency_ms":1680,"time_to_first_token_ms":258,"total_runtime_ms":1424,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0154,"confidence_interval":0.9348,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8881,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1204,"p95_latency_ms":1471,"p99_latency_ms":1578,"time_to_first_token_ms":253,"total_runtime_ms":1337,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0129,"confidence_interval":0.9376,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9831,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1248,"p95_latency_ms":1525,"p99_latency_ms":1636,"time_to_first_token_ms":237,"total_runtime_ms":1387,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0231,"confidence_interval":0.9417,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9278,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1189,"p95_latency_ms":1453,"p99_latency_ms":1559,"time_to_first_token_ms":251,"total_runtime_ms":1321,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.028,"confidence_interval":0.9302,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5509,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1237,"p95_latency_ms":1512,"p99_latency_ms":1622,"time_to_first_token_ms":238,"total_runtime_ms":1374,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0265,"confidence_interval":0.9406,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7366,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":692,"p95_latency_ms":846,"p99_latency_ms":908,"time_to_first_token_ms":202,"total_runtime_ms":769,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0154,"confidence_interval":0.9323,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5765,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":713,"p95_latency_ms":871,"p99_latency_ms":935,"time_to_first_token_ms":193,"total_runtime_ms":792,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0129,"confidence_interval":0.9499,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8178,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":758,"p95_latency_ms":927,"p99_latency_ms":994,"time_to_first_token_ms":206,"total_runtime_ms":842,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0229,"confidence_interval":0.9317,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7944,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":750,"p95_latency_ms":916,"p99_latency_ms":983,"time_to_first_token_ms":201,"total_runtime_ms":833,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0197,"confidence_interval":0.9449,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5004,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":205,"total_runtime_ms":849,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0216,"confidence_interval":0.9444,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0289,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":743,"p95_latency_ms":908,"p99_latency_ms":974,"time_to_first_token_ms":208,"total_runtime_ms":826,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0296,"confidence_interval":0.9419,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6454,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":726,"p95_latency_ms":887,"p99_latency_ms":951,"time_to_first_token_ms":222,"total_runtime_ms":806,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0169,"confidence_interval":0.9334,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3532,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":701,"p95_latency_ms":857,"p99_latency_ms":919,"time_to_first_token_ms":202,"total_runtime_ms":779,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0228,"confidence_interval":0.9349,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7278,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":712,"p95_latency_ms":870,"p99_latency_ms":933,"time_to_first_token_ms":203,"total_runtime_ms":791,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0166,"confidence_interval":0.9323,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7439,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":706,"p95_latency_ms":863,"p99_latency_ms":926,"time_to_first_token_ms":212,"total_runtime_ms":784,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.028,"confidence_interval":0.9404,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9454,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2479,"p95_latency_ms":3030,"p99_latency_ms":3251,"time_to_first_token_ms":422,"total_runtime_ms":2755,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0107,"confidence_interval":0.9462,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2936,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2453,"p95_latency_ms":2998,"p99_latency_ms":3216,"time_to_first_token_ms":418,"total_runtime_ms":2725,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0164,"confidence_interval":0.9434,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.975,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2398,"p95_latency_ms":2931,"p99_latency_ms":3144,"time_to_first_token_ms":416,"total_runtime_ms":2665,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0152,"confidence_interval":0.9347,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9008,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2419,"p95_latency_ms":2956,"p99_latency_ms":3171,"time_to_first_token_ms":406,"total_runtime_ms":2687,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0222,"confidence_interval":0.9337,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5127,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2482,"p95_latency_ms":3034,"p99_latency_ms":3254,"time_to_first_token_ms":424,"total_runtime_ms":2758,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0271,"confidence_interval":0.9476,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.519,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2421,"p95_latency_ms":2959,"p99_latency_ms":3175,"time_to_first_token_ms":421,"total_runtime_ms":2690,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0282,"confidence_interval":0.9466,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":8.609,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2381,"p95_latency_ms":2910,"p99_latency_ms":3121,"time_to_first_token_ms":433,"total_runtime_ms":2645,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0206,"confidence_interval":0.9364,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8926,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2447,"p95_latency_ms":2990,"p99_latency_ms":3208,"time_to_first_token_ms":433,"total_runtime_ms":2718,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0221,"confidence_interval":0.9371,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1205,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2441,"p95_latency_ms":2984,"p99_latency_ms":3201,"time_to_first_token_ms":413,"total_runtime_ms":2713,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0143,"confidence_interval":0.9335,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5014,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2458,"p95_latency_ms":3004,"p99_latency_ms":3223,"time_to_first_token_ms":417,"total_runtime_ms":2731,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0143,"confidence_interval":0.9464,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2492,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1415,"p95_latency_ms":1729,"p99_latency_ms":1855,"time_to_first_token_ms":394,"total_runtime_ms":1572,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.018,"confidence_interval":0.9425,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7448,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1454,"p95_latency_ms":1778,"p99_latency_ms":1907,"time_to_first_token_ms":384,"total_runtime_ms":1616,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0147,"confidence_interval":0.9458,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5167,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1381,"p95_latency_ms":1687,"p99_latency_ms":1810,"time_to_first_token_ms":391,"total_runtime_ms":1534,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0227,"confidence_interval":0.9319,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2726,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1472,"p95_latency_ms":1799,"p99_latency_ms":1930,"time_to_first_token_ms":391,"total_runtime_ms":1635,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.015,"confidence_interval":0.94,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0597,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1459,"p95_latency_ms":1783,"p99_latency_ms":1913,"time_to_first_token_ms":380,"total_runtime_ms":1621,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0188,"confidence_interval":0.9335,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7638,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1416,"p95_latency_ms":1731,"p99_latency_ms":1856,"time_to_first_token_ms":377,"total_runtime_ms":1573,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0218,"confidence_interval":0.9326,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5673,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1416,"p95_latency_ms":1731,"p99_latency_ms":1856,"time_to_first_token_ms":364,"total_runtime_ms":1573,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0256,"confidence_interval":0.9494,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.3036,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1420,"p95_latency_ms":1735,"p99_latency_ms":1861,"time_to_first_token_ms":366,"total_runtime_ms":1577,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0157,"confidence_interval":0.9469,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2866,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1412,"p95_latency_ms":1725,"p99_latency_ms":1851,"time_to_first_token_ms":388,"total_runtime_ms":1568,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0186,"confidence_interval":0.9334,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2669,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1450,"p95_latency_ms":1772,"p99_latency_ms":1901,"time_to_first_token_ms":373,"total_runtime_ms":1611,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0129,"confidence_interval":0.9384,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.3133,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3587,"p95_latency_ms":4384,"p99_latency_ms":4703,"time_to_first_token_ms":575,"total_runtime_ms":3986,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0204,"confidence_interval":0.9414,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6702,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3600,"p95_latency_ms":4400,"p99_latency_ms":4720,"time_to_first_token_ms":601,"total_runtime_ms":4000,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0164,"confidence_interval":0.9434,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0791,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3600,"p95_latency_ms":4400,"p99_latency_ms":4720,"time_to_first_token_ms":606,"total_runtime_ms":4000,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.018,"confidence_interval":0.9343,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9228,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3618,"p95_latency_ms":4421,"p99_latency_ms":4743,"time_to_first_token_ms":583,"total_runtime_ms":4019,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.9488,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3822,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3589,"p95_latency_ms":4387,"p99_latency_ms":4706,"time_to_first_token_ms":585,"total_runtime_ms":3988,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0268,"confidence_interval":0.9405,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2363,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3626,"p95_latency_ms":4431,"p99_latency_ms":4754,"time_to_first_token_ms":577,"total_runtime_ms":4028,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0194,"confidence_interval":0.947,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9989,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3576,"p95_latency_ms":4371,"p99_latency_ms":4689,"time_to_first_token_ms":595,"total_runtime_ms":3973,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.933,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2819,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3559,"p95_latency_ms":4350,"p99_latency_ms":4667,"time_to_first_token_ms":592,"total_runtime_ms":3955,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0119,"confidence_interval":0.93,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5052,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3579,"p95_latency_ms":4374,"p99_latency_ms":4692,"time_to_first_token_ms":575,"total_runtime_ms":3976,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.026,"confidence_interval":0.9492,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0249,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":3531,"p95_latency_ms":4315,"p99_latency_ms":4629,"time_to_first_token_ms":603,"total_runtime_ms":3923,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0178,"confidence_interval":0.9344,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3901,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2031,"p95_latency_ms":2482,"p99_latency_ms":2663,"time_to_first_token_ms":555,"total_runtime_ms":2256,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0159,"confidence_interval":0.9479,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8517,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2031,"p95_latency_ms":2483,"p99_latency_ms":2663,"time_to_first_token_ms":535,"total_runtime_ms":2257,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0158,"confidence_interval":0.9304,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9717,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2044,"p95_latency_ms":2498,"p99_latency_ms":2680,"time_to_first_token_ms":531,"total_runtime_ms":2271,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0141,"confidence_interval":0.9497,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5239,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2105,"p95_latency_ms":2572,"p99_latency_ms":2759,"time_to_first_token_ms":556,"total_runtime_ms":2339,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0216,"confidence_interval":0.9457,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.807,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2051,"p95_latency_ms":2506,"p99_latency_ms":2688,"time_to_first_token_ms":557,"total_runtime_ms":2278,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0102,"confidence_interval":0.9459,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5749,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2024,"p95_latency_ms":2474,"p99_latency_ms":2653,"time_to_first_token_ms":540,"total_runtime_ms":2249,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0118,"confidence_interval":0.9346,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7688,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2074,"p95_latency_ms":2535,"p99_latency_ms":2719,"time_to_first_token_ms":544,"total_runtime_ms":2304,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0126,"confidence_interval":0.9497,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6526,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2037,"p95_latency_ms":2490,"p99_latency_ms":2671,"time_to_first_token_ms":540,"total_runtime_ms":2263,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0293,"confidence_interval":0.9417,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6528,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2033,"p95_latency_ms":2484,"p99_latency_ms":2665,"time_to_first_token_ms":532,"total_runtime_ms":2258,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0262,"confidence_interval":0.9363,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5047,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|low|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2053,"p95_latency_ms":2509,"p99_latency_ms":2691,"time_to_first_token_ms":528,"total_runtime_ms":2281,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0143,"confidence_interval":0.9407,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7492,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":663,"p95_latency_ms":810,"p99_latency_ms":869,"time_to_first_token_ms":167,"total_runtime_ms":737,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0249,"confidence_interval":0.9393,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9359,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":746,"p95_latency_ms":911,"p99_latency_ms":978,"time_to_first_token_ms":160,"total_runtime_ms":829,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0257,"confidence_interval":0.9446,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8361,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":760,"p95_latency_ms":928,"p99_latency_ms":996,"time_to_first_token_ms":165,"total_runtime_ms":844,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0159,"confidence_interval":0.9311,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3519,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":718,"p95_latency_ms":878,"p99_latency_ms":942,"time_to_first_token_ms":183,"total_runtime_ms":798,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0104,"confidence_interval":0.9386,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0186,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":767,"p95_latency_ms":937,"p99_latency_ms":1006,"time_to_first_token_ms":183,"total_runtime_ms":852,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0109,"confidence_interval":0.935,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.824,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":740,"p95_latency_ms":904,"p99_latency_ms":970,"time_to_first_token_ms":169,"total_runtime_ms":822,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.029,"confidence_interval":0.9333,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9626,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":732,"p95_latency_ms":894,"p99_latency_ms":960,"time_to_first_token_ms":190,"total_runtime_ms":813,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0218,"confidence_interval":0.9371,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6741,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":733,"p95_latency_ms":896,"p99_latency_ms":962,"time_to_first_token_ms":161,"total_runtime_ms":815,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0263,"confidence_interval":0.9434,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5755,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":684,"p95_latency_ms":836,"p99_latency_ms":897,"time_to_first_token_ms":182,"total_runtime_ms":760,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0292,"confidence_interval":0.9482,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2751,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":176,"total_runtime_ms":764,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9335,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1218,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":555,"time_to_first_token_ms":144,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0229,"confidence_interval":0.9405,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4065,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":127,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0156,"confidence_interval":0.9333,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5304,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":507,"time_to_first_token_ms":138,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0138,"confidence_interval":0.9437,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8218,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":401,"p95_latency_ms":490,"p99_latency_ms":526,"time_to_first_token_ms":142,"total_runtime_ms":446,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0286,"confidence_interval":0.9398,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7894,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":432,"p95_latency_ms":528,"p99_latency_ms":567,"time_to_first_token_ms":143,"total_runtime_ms":480,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0146,"confidence_interval":0.9411,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.693,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":470,"p95_latency_ms":574,"p99_latency_ms":616,"time_to_first_token_ms":116,"total_runtime_ms":522,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0163,"confidence_interval":0.9416,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6703,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":453,"p95_latency_ms":554,"p99_latency_ms":594,"time_to_first_token_ms":118,"total_runtime_ms":504,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0113,"confidence_interval":0.9432,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3174,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":417,"p95_latency_ms":509,"p99_latency_ms":547,"time_to_first_token_ms":124,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0282,"confidence_interval":0.9463,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1795,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":434,"p95_latency_ms":530,"p99_latency_ms":569,"time_to_first_token_ms":143,"total_runtime_ms":482,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.017,"confidence_interval":0.9382,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5625,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":419,"p95_latency_ms":512,"p99_latency_ms":549,"time_to_first_token_ms":121,"total_runtime_ms":465,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0266,"confidence_interval":0.9413,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1806,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":898,"p95_latency_ms":1098,"p99_latency_ms":1178,"time_to_first_token_ms":214,"total_runtime_ms":998,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0232,"confidence_interval":0.9423,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6314,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":912,"p95_latency_ms":1114,"p99_latency_ms":1195,"time_to_first_token_ms":217,"total_runtime_ms":1013,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0126,"confidence_interval":0.9411,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.275,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":994,"p95_latency_ms":1215,"p99_latency_ms":1304,"time_to_first_token_ms":202,"total_runtime_ms":1105,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0115,"confidence_interval":0.9382,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8836,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":940,"p95_latency_ms":1148,"p99_latency_ms":1232,"time_to_first_token_ms":217,"total_runtime_ms":1044,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0212,"confidence_interval":0.9429,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1523,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":994,"p95_latency_ms":1215,"p99_latency_ms":1303,"time_to_first_token_ms":206,"total_runtime_ms":1104,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0153,"confidence_interval":0.9418,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0441,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":977,"p95_latency_ms":1194,"p99_latency_ms":1280,"time_to_first_token_ms":213,"total_runtime_ms":1085,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0178,"confidence_interval":0.9352,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.195,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":925,"p95_latency_ms":1131,"p99_latency_ms":1213,"time_to_first_token_ms":190,"total_runtime_ms":1028,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0272,"confidence_interval":0.9341,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4666,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":959,"p95_latency_ms":1173,"p99_latency_ms":1258,"time_to_first_token_ms":201,"total_runtime_ms":1066,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0193,"confidence_interval":0.9455,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7444,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":967,"p95_latency_ms":1181,"p99_latency_ms":1267,"time_to_first_token_ms":217,"total_runtime_ms":1074,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0105,"confidence_interval":0.9487,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1007,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":894,"p95_latency_ms":1092,"p99_latency_ms":1172,"time_to_first_token_ms":217,"total_runtime_ms":993,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0246,"confidence_interval":0.9316,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7305,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":537,"p95_latency_ms":656,"p99_latency_ms":704,"time_to_first_token_ms":146,"total_runtime_ms":596,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0248,"confidence_interval":0.9334,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6521,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":170,"total_runtime_ms":586,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0264,"confidence_interval":0.9387,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3285,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":617,"p95_latency_ms":754,"p99_latency_ms":809,"time_to_first_token_ms":174,"total_runtime_ms":685,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0157,"confidence_interval":0.9373,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8045,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":542,"p95_latency_ms":663,"p99_latency_ms":711,"time_to_first_token_ms":145,"total_runtime_ms":603,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0183,"confidence_interval":0.9421,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3955,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":610,"p95_latency_ms":746,"p99_latency_ms":800,"time_to_first_token_ms":166,"total_runtime_ms":678,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0109,"confidence_interval":0.9342,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6935,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":541,"p95_latency_ms":661,"p99_latency_ms":709,"time_to_first_token_ms":168,"total_runtime_ms":601,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0207,"confidence_interval":0.9351,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7907,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":786,"time_to_first_token_ms":153,"total_runtime_ms":666,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0253,"confidence_interval":0.9468,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1219,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":570,"p95_latency_ms":696,"p99_latency_ms":747,"time_to_first_token_ms":172,"total_runtime_ms":633,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.025,"confidence_interval":0.9419,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.461,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":773,"time_to_first_token_ms":165,"total_runtime_ms":655,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0166,"confidence_interval":0.9355,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7823,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":533,"p95_latency_ms":652,"p99_latency_ms":699,"time_to_first_token_ms":149,"total_runtime_ms":593,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0186,"confidence_interval":0.9461,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1255,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1879,"p95_latency_ms":2296,"p99_latency_ms":2463,"time_to_first_token_ms":345,"total_runtime_ms":2087,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0103,"confidence_interval":0.942,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5999,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1831,"p95_latency_ms":2238,"p99_latency_ms":2401,"time_to_first_token_ms":332,"total_runtime_ms":2035,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0192,"confidence_interval":0.9351,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2847,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1801,"p95_latency_ms":2202,"p99_latency_ms":2362,"time_to_first_token_ms":332,"total_runtime_ms":2002,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0149,"confidence_interval":0.9393,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3284,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1782,"p95_latency_ms":2178,"p99_latency_ms":2336,"time_to_first_token_ms":340,"total_runtime_ms":1980,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.018,"confidence_interval":0.9317,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0333,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1824,"p95_latency_ms":2229,"p99_latency_ms":2391,"time_to_first_token_ms":348,"total_runtime_ms":2027,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0209,"confidence_interval":0.9322,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6932,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1878,"p95_latency_ms":2295,"p99_latency_ms":2462,"time_to_first_token_ms":337,"total_runtime_ms":2087,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0161,"confidence_interval":0.9387,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7863,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1851,"p95_latency_ms":2263,"p99_latency_ms":2427,"time_to_first_token_ms":345,"total_runtime_ms":2057,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0219,"confidence_interval":0.9473,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2328,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1815,"p95_latency_ms":2219,"p99_latency_ms":2380,"time_to_first_token_ms":341,"total_runtime_ms":2017,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0201,"confidence_interval":0.9484,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5103,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1819,"p95_latency_ms":2223,"p99_latency_ms":2385,"time_to_first_token_ms":348,"total_runtime_ms":2021,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0231,"confidence_interval":0.9331,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1162,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":1860,"p95_latency_ms":2274,"p99_latency_ms":2439,"time_to_first_token_ms":351,"total_runtime_ms":2067,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0139,"confidence_interval":0.9391,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3896,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1031,"p95_latency_ms":1261,"p99_latency_ms":1352,"time_to_first_token_ms":276,"total_runtime_ms":1146,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0173,"confidence_interval":0.9374,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1218,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1042,"p95_latency_ms":1273,"p99_latency_ms":1366,"time_to_first_token_ms":300,"total_runtime_ms":1158,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0185,"confidence_interval":0.9384,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2189,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1344,"time_to_first_token_ms":285,"total_runtime_ms":1139,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0158,"confidence_interval":0.9413,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3324,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1069,"p95_latency_ms":1307,"p99_latency_ms":1402,"time_to_first_token_ms":308,"total_runtime_ms":1188,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0296,"confidence_interval":0.9358,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9746,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1118,"p95_latency_ms":1366,"p99_latency_ms":1465,"time_to_first_token_ms":280,"total_runtime_ms":1242,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0158,"confidence_interval":0.9456,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9395,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1041,"p95_latency_ms":1273,"p99_latency_ms":1365,"time_to_first_token_ms":302,"total_runtime_ms":1157,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.028,"confidence_interval":0.9493,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0089,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1118,"p95_latency_ms":1366,"p99_latency_ms":1465,"time_to_first_token_ms":303,"total_runtime_ms":1242,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0253,"confidence_interval":0.9375,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5287,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1077,"p95_latency_ms":1316,"p99_latency_ms":1412,"time_to_first_token_ms":276,"total_runtime_ms":1196,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.01,"confidence_interval":0.9419,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7861,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1112,"p95_latency_ms":1359,"p99_latency_ms":1458,"time_to_first_token_ms":295,"total_runtime_ms":1235,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0166,"confidence_interval":0.9455,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3075,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1107,"p95_latency_ms":1353,"p99_latency_ms":1451,"time_to_first_token_ms":276,"total_runtime_ms":1230,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0299,"confidence_interval":0.9374,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.57,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2683,"p95_latency_ms":3279,"p99_latency_ms":3518,"time_to_first_token_ms":466,"total_runtime_ms":2981,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0144,"confidence_interval":0.9414,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7115,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2642,"p95_latency_ms":3229,"p99_latency_ms":3463,"time_to_first_token_ms":474,"total_runtime_ms":2935,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0203,"confidence_interval":0.9361,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3162,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2738,"p95_latency_ms":3346,"p99_latency_ms":3589,"time_to_first_token_ms":456,"total_runtime_ms":3042,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0298,"confidence_interval":0.9498,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":11.254,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2642,"p95_latency_ms":3230,"p99_latency_ms":3464,"time_to_first_token_ms":460,"total_runtime_ms":2936,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.9329,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0876,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2672,"p95_latency_ms":3266,"p99_latency_ms":3503,"time_to_first_token_ms":475,"total_runtime_ms":2969,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0179,"confidence_interval":0.9323,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7808,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2723,"p95_latency_ms":3328,"p99_latency_ms":3570,"time_to_first_token_ms":478,"total_runtime_ms":3026,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0271,"confidence_interval":0.9485,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4323,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2688,"p95_latency_ms":3286,"p99_latency_ms":3525,"time_to_first_token_ms":473,"total_runtime_ms":2987,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0202,"confidence_interval":0.9489,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0942,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2647,"p95_latency_ms":3235,"p99_latency_ms":3470,"time_to_first_token_ms":451,"total_runtime_ms":2941,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0177,"confidence_interval":0.9486,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8067,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2648,"p95_latency_ms":3236,"p99_latency_ms":3472,"time_to_first_token_ms":470,"total_runtime_ms":2942,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0231,"confidence_interval":0.9378,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3404,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":2697,"p95_latency_ms":3296,"p99_latency_ms":3536,"time_to_first_token_ms":447,"total_runtime_ms":2997,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.014,"confidence_interval":0.9463,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8191,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1524,"p95_latency_ms":1863,"p99_latency_ms":1999,"time_to_first_token_ms":416,"total_runtime_ms":1694,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.9431,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.282,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1548,"p95_latency_ms":1892,"p99_latency_ms":2030,"time_to_first_token_ms":413,"total_runtime_ms":1720,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0274,"confidence_interval":0.9475,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6959,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1593,"p95_latency_ms":1947,"p99_latency_ms":2089,"time_to_first_token_ms":409,"total_runtime_ms":1770,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0119,"confidence_interval":0.9454,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1278,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1529,"p95_latency_ms":1869,"p99_latency_ms":2005,"time_to_first_token_ms":406,"total_runtime_ms":1699,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0102,"confidence_interval":0.9387,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9617,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1545,"p95_latency_ms":1888,"p99_latency_ms":2025,"time_to_first_token_ms":413,"total_runtime_ms":1716,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.018,"confidence_interval":0.9318,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9368,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1599,"p95_latency_ms":1954,"p99_latency_ms":2096,"time_to_first_token_ms":414,"total_runtime_ms":1776,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0117,"confidence_interval":0.9336,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0127,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1576,"p95_latency_ms":1927,"p99_latency_ms":2067,"time_to_first_token_ms":423,"total_runtime_ms":1751,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0101,"confidence_interval":0.9326,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.125,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1558,"p95_latency_ms":1904,"p99_latency_ms":2042,"time_to_first_token_ms":422,"total_runtime_ms":1731,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0127,"confidence_interval":0.9414,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":11.175,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1576,"p95_latency_ms":1927,"p99_latency_ms":2067,"time_to_first_token_ms":430,"total_runtime_ms":1751,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.029,"confidence_interval":0.949,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9995,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":1600,"p95_latency_ms":1956,"p99_latency_ms":2098,"time_to_first_token_ms":421,"total_runtime_ms":1778,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0204,"confidence_interval":0.9449,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5576,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":519,"p95_latency_ms":635,"p99_latency_ms":681,"time_to_first_token_ms":145,"total_runtime_ms":577,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0186,"confidence_interval":0.9393,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3123,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":597,"p95_latency_ms":729,"p99_latency_ms":782,"time_to_first_token_ms":150,"total_runtime_ms":663,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0141,"confidence_interval":0.9476,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0409,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":534,"p95_latency_ms":653,"p99_latency_ms":700,"time_to_first_token_ms":164,"total_runtime_ms":594,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0288,"confidence_interval":0.9311,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6891,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":536,"p95_latency_ms":655,"p99_latency_ms":703,"time_to_first_token_ms":141,"total_runtime_ms":596,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0121,"confidence_interval":0.9462,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1903,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":138,"total_runtime_ms":591,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0233,"confidence_interval":0.9468,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9832,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":545,"p95_latency_ms":666,"p99_latency_ms":715,"time_to_first_token_ms":147,"total_runtime_ms":606,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0143,"confidence_interval":0.9314,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.923,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":785,"time_to_first_token_ms":152,"total_runtime_ms":665,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0204,"confidence_interval":0.9328,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5938,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":148,"total_runtime_ms":575,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0291,"confidence_interval":0.9378,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8895,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":581,"p95_latency_ms":710,"p99_latency_ms":762,"time_to_first_token_ms":153,"total_runtime_ms":646,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0237,"confidence_interval":0.9379,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.653,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":544,"p95_latency_ms":665,"p99_latency_ms":713,"time_to_first_token_ms":154,"total_runtime_ms":604,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.011,"confidence_interval":0.9349,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3536,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":344,"p95_latency_ms":420,"p99_latency_ms":451,"time_to_first_token_ms":112,"total_runtime_ms":382,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0134,"confidence_interval":0.9379,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9017,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":506,"time_to_first_token_ms":116,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.012,"confidence_interval":0.9416,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.145,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":338,"p95_latency_ms":414,"p99_latency_ms":444,"time_to_first_token_ms":115,"total_runtime_ms":376,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0134,"confidence_interval":0.9421,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0117,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":346,"p95_latency_ms":423,"p99_latency_ms":454,"time_to_first_token_ms":110,"total_runtime_ms":385,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0186,"confidence_interval":0.9384,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2733,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":354,"p95_latency_ms":433,"p99_latency_ms":464,"time_to_first_token_ms":91,"total_runtime_ms":393,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0227,"confidence_interval":0.9479,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1519,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":333,"p95_latency_ms":407,"p99_latency_ms":437,"time_to_first_token_ms":94,"total_runtime_ms":370,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0227,"confidence_interval":0.9491,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9278,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":313,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":87,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0111,"confidence_interval":0.9431,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3762,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":310,"p95_latency_ms":378,"p99_latency_ms":406,"time_to_first_token_ms":89,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.024,"confidence_interval":0.9495,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2006,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":303,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":93,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0205,"confidence_interval":0.931,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5552,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":360,"p95_latency_ms":440,"p99_latency_ms":472,"time_to_first_token_ms":107,"total_runtime_ms":400,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0286,"confidence_interval":0.9313,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3894,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":771,"p95_latency_ms":942,"p99_latency_ms":1011,"time_to_first_token_ms":167,"total_runtime_ms":856,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0188,"confidence_interval":0.946,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0901,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":670,"p95_latency_ms":819,"p99_latency_ms":879,"time_to_first_token_ms":183,"total_runtime_ms":745,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0211,"confidence_interval":0.9476,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4303,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":751,"p95_latency_ms":918,"p99_latency_ms":985,"time_to_first_token_ms":187,"total_runtime_ms":835,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0244,"confidence_interval":0.9479,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0306,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":694,"p95_latency_ms":848,"p99_latency_ms":910,"time_to_first_token_ms":162,"total_runtime_ms":771,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0177,"confidence_interval":0.9473,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6212,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":696,"p95_latency_ms":851,"p99_latency_ms":913,"time_to_first_token_ms":186,"total_runtime_ms":774,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0224,"confidence_interval":0.9375,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8198,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":747,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":171,"total_runtime_ms":831,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9385,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2503,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":744,"p95_latency_ms":910,"p99_latency_ms":976,"time_to_first_token_ms":168,"total_runtime_ms":827,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0181,"confidence_interval":0.943,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6713,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":760,"p95_latency_ms":929,"p99_latency_ms":996,"time_to_first_token_ms":179,"total_runtime_ms":844,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0243,"confidence_interval":0.9381,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.915,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":717,"p95_latency_ms":876,"p99_latency_ms":940,"time_to_first_token_ms":160,"total_runtime_ms":797,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0239,"confidence_interval":0.9409,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6008,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":684,"p95_latency_ms":836,"p99_latency_ms":897,"time_to_first_token_ms":167,"total_runtime_ms":760,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0235,"confidence_interval":0.9467,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4221,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":413,"p95_latency_ms":505,"p99_latency_ms":542,"time_to_first_token_ms":132,"total_runtime_ms":459,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0165,"confidence_interval":0.9397,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8071,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":416,"p95_latency_ms":509,"p99_latency_ms":546,"time_to_first_token_ms":122,"total_runtime_ms":462,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0191,"confidence_interval":0.9392,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0277,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":142,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0122,"confidence_interval":0.9339,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7733,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":414,"p95_latency_ms":506,"p99_latency_ms":543,"time_to_first_token_ms":139,"total_runtime_ms":460,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0167,"confidence_interval":0.9376,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.831,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":126,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0275,"confidence_interval":0.9398,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3102,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":416,"p95_latency_ms":508,"p99_latency_ms":545,"time_to_first_token_ms":143,"total_runtime_ms":462,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0144,"confidence_interval":0.9356,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5679,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":121,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0153,"confidence_interval":0.9312,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7193,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":483,"p95_latency_ms":590,"p99_latency_ms":633,"time_to_first_token_ms":113,"total_runtime_ms":537,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0166,"confidence_interval":0.9309,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1411,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":445,"p95_latency_ms":544,"p99_latency_ms":584,"time_to_first_token_ms":132,"total_runtime_ms":495,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0171,"confidence_interval":0.9457,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8036,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":445,"p95_latency_ms":544,"p99_latency_ms":584,"time_to_first_token_ms":127,"total_runtime_ms":495,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0131,"confidence_interval":0.9358,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2247,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1357,"p95_latency_ms":1659,"p99_latency_ms":1779,"time_to_first_token_ms":279,"total_runtime_ms":1508,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0249,"confidence_interval":0.9402,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":9.3055,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1418,"p95_latency_ms":1733,"p99_latency_ms":1860,"time_to_first_token_ms":256,"total_runtime_ms":1576,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0158,"confidence_interval":0.9421,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":9.3016,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1368,"p95_latency_ms":1672,"p99_latency_ms":1794,"time_to_first_token_ms":282,"total_runtime_ms":1520,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0241,"confidence_interval":0.9352,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7757,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1373,"p95_latency_ms":1678,"p99_latency_ms":1801,"time_to_first_token_ms":274,"total_runtime_ms":1526,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0198,"confidence_interval":0.9419,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2156,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1394,"p95_latency_ms":1704,"p99_latency_ms":1828,"time_to_first_token_ms":265,"total_runtime_ms":1549,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0204,"confidence_interval":0.9459,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6102,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1391,"p95_latency_ms":1700,"p99_latency_ms":1824,"time_to_first_token_ms":262,"total_runtime_ms":1545,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0285,"confidence_interval":0.9413,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6933,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1376,"p95_latency_ms":1682,"p99_latency_ms":1804,"time_to_first_token_ms":277,"total_runtime_ms":1529,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9478,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8239,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1348,"p95_latency_ms":1647,"p99_latency_ms":1767,"time_to_first_token_ms":265,"total_runtime_ms":1497,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0166,"confidence_interval":0.9358,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7801,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1427,"p95_latency_ms":1744,"p99_latency_ms":1871,"time_to_first_token_ms":269,"total_runtime_ms":1586,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0187,"confidence_interval":0.9341,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7192,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1415,"p95_latency_ms":1730,"p99_latency_ms":1855,"time_to_first_token_ms":261,"total_runtime_ms":1572,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0219,"confidence_interval":0.9362,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9525,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":842,"p95_latency_ms":1029,"p99_latency_ms":1104,"time_to_first_token_ms":243,"total_runtime_ms":936,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0281,"confidence_interval":0.9348,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2619,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":811,"p95_latency_ms":991,"p99_latency_ms":1063,"time_to_first_token_ms":223,"total_runtime_ms":901,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0127,"confidence_interval":0.9419,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3991,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":804,"p95_latency_ms":982,"p99_latency_ms":1054,"time_to_first_token_ms":234,"total_runtime_ms":893,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0162,"confidence_interval":0.9385,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":9.3132,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":855,"p95_latency_ms":1045,"p99_latency_ms":1121,"time_to_first_token_ms":227,"total_runtime_ms":950,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0149,"confidence_interval":0.9496,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":9.261,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":874,"p95_latency_ms":1068,"p99_latency_ms":1145,"time_to_first_token_ms":239,"total_runtime_ms":971,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0131,"confidence_interval":0.9302,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.533,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":793,"p95_latency_ms":969,"p99_latency_ms":1040,"time_to_first_token_ms":238,"total_runtime_ms":881,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0159,"confidence_interval":0.9319,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2449,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":822,"p95_latency_ms":1005,"p99_latency_ms":1078,"time_to_first_token_ms":231,"total_runtime_ms":913,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0233,"confidence_interval":0.9411,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4822,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":847,"p95_latency_ms":1035,"p99_latency_ms":1110,"time_to_first_token_ms":213,"total_runtime_ms":941,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0208,"confidence_interval":0.9468,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8303,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":841,"p95_latency_ms":1027,"p99_latency_ms":1102,"time_to_first_token_ms":227,"total_runtime_ms":934,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0235,"confidence_interval":0.9488,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2945,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":862,"p95_latency_ms":1054,"p99_latency_ms":1130,"time_to_first_token_ms":226,"total_runtime_ms":958,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.9468,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4395,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2030,"p95_latency_ms":2482,"p99_latency_ms":2662,"time_to_first_token_ms":350,"total_runtime_ms":2256,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0101,"confidence_interval":0.9484,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0808,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2015,"p95_latency_ms":2463,"p99_latency_ms":2642,"time_to_first_token_ms":349,"total_runtime_ms":2239,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0298,"confidence_interval":0.946,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2338,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2043,"p95_latency_ms":2497,"p99_latency_ms":2679,"time_to_first_token_ms":358,"total_runtime_ms":2270,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0269,"confidence_interval":0.9492,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8248,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2008,"p95_latency_ms":2454,"p99_latency_ms":2633,"time_to_first_token_ms":373,"total_runtime_ms":2231,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0198,"confidence_interval":0.9337,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2651,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2034,"p95_latency_ms":2486,"p99_latency_ms":2667,"time_to_first_token_ms":378,"total_runtime_ms":2260,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0174,"confidence_interval":0.9333,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5691,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2057,"p95_latency_ms":2514,"p99_latency_ms":2696,"time_to_first_token_ms":371,"total_runtime_ms":2285,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0214,"confidence_interval":0.9349,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9149,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2029,"p95_latency_ms":2479,"p99_latency_ms":2660,"time_to_first_token_ms":363,"total_runtime_ms":2254,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0138,"confidence_interval":0.9325,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5287,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2018,"p95_latency_ms":2466,"p99_latency_ms":2646,"time_to_first_token_ms":367,"total_runtime_ms":2242,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0288,"confidence_interval":0.937,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9274,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2084,"p95_latency_ms":2547,"p99_latency_ms":2732,"time_to_first_token_ms":375,"total_runtime_ms":2315,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.013,"confidence_interval":0.9401,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2416,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2033,"p95_latency_ms":2485,"p99_latency_ms":2666,"time_to_first_token_ms":374,"total_runtime_ms":2259,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0187,"confidence_interval":0.9435,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8707,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1191,"p95_latency_ms":1456,"p99_latency_ms":1561,"time_to_first_token_ms":307,"total_runtime_ms":1323,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0165,"confidence_interval":0.9302,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2084,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1179,"p95_latency_ms":1441,"p99_latency_ms":1546,"time_to_first_token_ms":321,"total_runtime_ms":1310,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0176,"confidence_interval":0.9498,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2733,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1173,"p95_latency_ms":1434,"p99_latency_ms":1538,"time_to_first_token_ms":318,"total_runtime_ms":1303,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0242,"confidence_interval":0.9423,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6242,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1201,"p95_latency_ms":1468,"p99_latency_ms":1575,"time_to_first_token_ms":309,"total_runtime_ms":1334,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0135,"confidence_interval":0.9331,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7737,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1214,"p95_latency_ms":1483,"p99_latency_ms":1591,"time_to_first_token_ms":304,"total_runtime_ms":1348,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0217,"confidence_interval":0.9496,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3719,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1180,"p95_latency_ms":1442,"p99_latency_ms":1547,"time_to_first_token_ms":309,"total_runtime_ms":1311,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0193,"confidence_interval":0.9432,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8397,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1159,"p95_latency_ms":1416,"p99_latency_ms":1519,"time_to_first_token_ms":333,"total_runtime_ms":1287,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0117,"confidence_interval":0.9381,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6974,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1188,"p95_latency_ms":1452,"p99_latency_ms":1558,"time_to_first_token_ms":310,"total_runtime_ms":1320,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0259,"confidence_interval":0.9452,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0117,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1199,"p95_latency_ms":1465,"p99_latency_ms":1572,"time_to_first_token_ms":304,"total_runtime_ms":1332,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0238,"confidence_interval":0.9344,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8319,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1171,"p95_latency_ms":1431,"p99_latency_ms":1535,"time_to_first_token_ms":314,"total_runtime_ms":1301,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0273,"confidence_interval":0.9449,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9008,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":498,"p95_latency_ms":608,"p99_latency_ms":653,"time_to_first_token_ms":131,"total_runtime_ms":553,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9494,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5258,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":120,"total_runtime_ms":476,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0279,"confidence_interval":0.9468,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9992,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":416,"p95_latency_ms":508,"p99_latency_ms":545,"time_to_first_token_ms":130,"total_runtime_ms":462,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0211,"confidence_interval":0.9475,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2135,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":445,"p95_latency_ms":544,"p99_latency_ms":583,"time_to_first_token_ms":143,"total_runtime_ms":494,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0254,"confidence_interval":0.9398,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9789,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":474,"p95_latency_ms":580,"p99_latency_ms":622,"time_to_first_token_ms":123,"total_runtime_ms":527,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0288,"confidence_interval":0.9463,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6045,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":502,"p95_latency_ms":613,"p99_latency_ms":658,"time_to_first_token_ms":151,"total_runtime_ms":557,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0175,"confidence_interval":0.9318,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8293,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":509,"p95_latency_ms":622,"p99_latency_ms":668,"time_to_first_token_ms":120,"total_runtime_ms":566,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0177,"confidence_interval":0.935,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3983,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":124,"total_runtime_ms":475,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0248,"confidence_interval":0.9319,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7189,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":483,"p95_latency_ms":591,"p99_latency_ms":634,"time_to_first_token_ms":127,"total_runtime_ms":537,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0182,"confidence_interval":0.9429,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6562,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":462,"p95_latency_ms":564,"p99_latency_ms":605,"time_to_first_token_ms":155,"total_runtime_ms":513,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0261,"confidence_interval":0.9421,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8483,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":332,"p95_latency_ms":405,"p99_latency_ms":435,"time_to_first_token_ms":105,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0208,"confidence_interval":0.945,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7405,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":323,"p95_latency_ms":395,"p99_latency_ms":423,"time_to_first_token_ms":100,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0254,"confidence_interval":0.9337,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4255,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":327,"p95_latency_ms":399,"p99_latency_ms":428,"time_to_first_token_ms":94,"total_runtime_ms":363,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0296,"confidence_interval":0.9327,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8807,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":327,"p95_latency_ms":399,"p99_latency_ms":428,"time_to_first_token_ms":89,"total_runtime_ms":363,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0271,"confidence_interval":0.9406,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7012,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":298,"p95_latency_ms":364,"p99_latency_ms":390,"time_to_first_token_ms":101,"total_runtime_ms":331,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0122,"confidence_interval":0.9344,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5622,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":335,"p95_latency_ms":410,"p99_latency_ms":439,"time_to_first_token_ms":102,"total_runtime_ms":372,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0277,"confidence_interval":0.9489,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7434,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":291,"p95_latency_ms":355,"p99_latency_ms":381,"time_to_first_token_ms":80,"total_runtime_ms":323,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0152,"confidence_interval":0.9347,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8728,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":293,"p95_latency_ms":358,"p99_latency_ms":385,"time_to_first_token_ms":95,"total_runtime_ms":326,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.018,"confidence_interval":0.9394,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6624,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":329,"p95_latency_ms":402,"p99_latency_ms":431,"time_to_first_token_ms":103,"total_runtime_ms":365,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0155,"confidence_interval":0.9372,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5065,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":291,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":85,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0217,"confidence_interval":0.9471,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4439,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":843,"time_to_first_token_ms":161,"total_runtime_ms":715,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0269,"confidence_interval":0.9339,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1308,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":616,"p95_latency_ms":752,"p99_latency_ms":807,"time_to_first_token_ms":145,"total_runtime_ms":684,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0278,"confidence_interval":0.9383,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4875,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":603,"p95_latency_ms":737,"p99_latency_ms":790,"time_to_first_token_ms":152,"total_runtime_ms":670,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0296,"confidence_interval":0.9374,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9007,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":658,"p95_latency_ms":804,"p99_latency_ms":862,"time_to_first_token_ms":165,"total_runtime_ms":731,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0197,"confidence_interval":0.9436,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3593,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":587,"p95_latency_ms":717,"p99_latency_ms":770,"time_to_first_token_ms":143,"total_runtime_ms":652,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0228,"confidence_interval":0.9361,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9669,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":622,"p95_latency_ms":760,"p99_latency_ms":816,"time_to_first_token_ms":146,"total_runtime_ms":691,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0242,"confidence_interval":0.948,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6865,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":596,"p95_latency_ms":728,"p99_latency_ms":781,"time_to_first_token_ms":162,"total_runtime_ms":662,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0159,"confidence_interval":0.9478,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5775,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":802,"time_to_first_token_ms":167,"total_runtime_ms":680,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0122,"confidence_interval":0.9467,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7218,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":624,"p95_latency_ms":763,"p99_latency_ms":819,"time_to_first_token_ms":148,"total_runtime_ms":694,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0136,"confidence_interval":0.9337,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6494,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":601,"p95_latency_ms":735,"p99_latency_ms":788,"time_to_first_token_ms":143,"total_runtime_ms":668,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0145,"confidence_interval":0.9429,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4081,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":349,"p95_latency_ms":427,"p99_latency_ms":458,"time_to_first_token_ms":121,"total_runtime_ms":388,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0131,"confidence_interval":0.9488,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9609,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":407,"p95_latency_ms":497,"p99_latency_ms":533,"time_to_first_token_ms":121,"total_runtime_ms":452,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0226,"confidence_interval":0.9485,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5291,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":336,"p95_latency_ms":411,"p99_latency_ms":441,"time_to_first_token_ms":98,"total_runtime_ms":374,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0164,"confidence_interval":0.9484,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3013,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":346,"p95_latency_ms":423,"p99_latency_ms":453,"time_to_first_token_ms":122,"total_runtime_ms":384,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0199,"confidence_interval":0.9438,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8379,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":348,"p95_latency_ms":426,"p99_latency_ms":457,"time_to_first_token_ms":131,"total_runtime_ms":387,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0138,"confidence_interval":0.9403,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4493,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":334,"p95_latency_ms":409,"p99_latency_ms":438,"time_to_first_token_ms":122,"total_runtime_ms":372,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9374,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5044,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":422,"p95_latency_ms":515,"p99_latency_ms":553,"time_to_first_token_ms":115,"total_runtime_ms":468,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.016,"confidence_interval":0.9372,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.807,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":397,"p95_latency_ms":485,"p99_latency_ms":520,"time_to_first_token_ms":99,"total_runtime_ms":441,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0266,"confidence_interval":0.9341,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1971,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":345,"p95_latency_ms":421,"p99_latency_ms":452,"time_to_first_token_ms":119,"total_runtime_ms":383,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0298,"confidence_interval":0.9326,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8518,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":372,"p95_latency_ms":454,"p99_latency_ms":487,"time_to_first_token_ms":118,"total_runtime_ms":413,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0292,"confidence_interval":0.9371,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7739,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1193,"p95_latency_ms":1458,"p99_latency_ms":1564,"time_to_first_token_ms":245,"total_runtime_ms":1326,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.9464,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5545,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1173,"p95_latency_ms":1434,"p99_latency_ms":1538,"time_to_first_token_ms":256,"total_runtime_ms":1304,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0124,"confidence_interval":0.9342,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.421,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1212,"p95_latency_ms":1481,"p99_latency_ms":1589,"time_to_first_token_ms":228,"total_runtime_ms":1347,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0223,"confidence_interval":0.9349,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7497,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1182,"p95_latency_ms":1445,"p99_latency_ms":1550,"time_to_first_token_ms":223,"total_runtime_ms":1314,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0195,"confidence_interval":0.9465,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8521,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1151,"p95_latency_ms":1407,"p99_latency_ms":1509,"time_to_first_token_ms":231,"total_runtime_ms":1279,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0217,"confidence_interval":0.9445,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9094,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1210,"p95_latency_ms":1479,"p99_latency_ms":1586,"time_to_first_token_ms":239,"total_runtime_ms":1344,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.019,"confidence_interval":0.9469,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2509,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1191,"p95_latency_ms":1456,"p99_latency_ms":1562,"time_to_first_token_ms":243,"total_runtime_ms":1324,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0138,"confidence_interval":0.9419,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6648,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1180,"p95_latency_ms":1443,"p99_latency_ms":1548,"time_to_first_token_ms":250,"total_runtime_ms":1312,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0123,"confidence_interval":0.9327,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1083,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1116,"p95_latency_ms":1365,"p99_latency_ms":1464,"time_to_first_token_ms":240,"total_runtime_ms":1240,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0279,"confidence_interval":0.9416,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9596,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1131,"p95_latency_ms":1382,"p99_latency_ms":1483,"time_to_first_token_ms":240,"total_runtime_ms":1257,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0272,"confidence_interval":0.9462,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3624,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":641,"p95_latency_ms":783,"p99_latency_ms":840,"time_to_first_token_ms":211,"total_runtime_ms":712,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0217,"confidence_interval":0.9486,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3676,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":715,"p95_latency_ms":874,"p99_latency_ms":937,"time_to_first_token_ms":194,"total_runtime_ms":794,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.021,"confidence_interval":0.945,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3998,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":721,"p95_latency_ms":881,"p99_latency_ms":945,"time_to_first_token_ms":191,"total_runtime_ms":801,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0259,"confidence_interval":0.9319,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8067,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":642,"p95_latency_ms":785,"p99_latency_ms":842,"time_to_first_token_ms":204,"total_runtime_ms":714,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0287,"confidence_interval":0.9443,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":9.3166,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":705,"p95_latency_ms":862,"p99_latency_ms":925,"time_to_first_token_ms":204,"total_runtime_ms":783,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0239,"confidence_interval":0.9456,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":9.3114,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":657,"p95_latency_ms":803,"p99_latency_ms":862,"time_to_first_token_ms":200,"total_runtime_ms":730,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0174,"confidence_interval":0.9497,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1613,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":730,"p95_latency_ms":892,"p99_latency_ms":957,"time_to_first_token_ms":201,"total_runtime_ms":811,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0294,"confidence_interval":0.9361,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.435,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":734,"p95_latency_ms":897,"p99_latency_ms":962,"time_to_first_token_ms":208,"total_runtime_ms":815,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0185,"confidence_interval":0.937,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9451,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":747,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":195,"total_runtime_ms":830,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0288,"confidence_interval":0.9452,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7912,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":701,"p95_latency_ms":857,"p99_latency_ms":919,"time_to_first_token_ms":192,"total_runtime_ms":779,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.014,"confidence_interval":0.9342,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5335,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1666,"p95_latency_ms":2037,"p99_latency_ms":2185,"time_to_first_token_ms":302,"total_runtime_ms":1852,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0189,"confidence_interval":0.9469,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7536,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1724,"p95_latency_ms":2107,"p99_latency_ms":2261,"time_to_first_token_ms":313,"total_runtime_ms":1916,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0246,"confidence_interval":0.9373,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4881,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1752,"p95_latency_ms":2141,"p99_latency_ms":2297,"time_to_first_token_ms":301,"total_runtime_ms":1947,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0145,"confidence_interval":0.9327,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7852,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1683,"p95_latency_ms":2056,"p99_latency_ms":2206,"time_to_first_token_ms":327,"total_runtime_ms":1870,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0197,"confidence_interval":0.9319,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4657,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1736,"p95_latency_ms":2122,"p99_latency_ms":2276,"time_to_first_token_ms":312,"total_runtime_ms":1929,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0297,"confidence_interval":0.9334,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7999,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1735,"p95_latency_ms":2121,"p99_latency_ms":2275,"time_to_first_token_ms":305,"total_runtime_ms":1928,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0279,"confidence_interval":0.9317,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0708,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1696,"p95_latency_ms":2073,"p99_latency_ms":2223,"time_to_first_token_ms":332,"total_runtime_ms":1884,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0186,"confidence_interval":0.9323,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8747,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1674,"p95_latency_ms":2046,"p99_latency_ms":2195,"time_to_first_token_ms":310,"total_runtime_ms":1860,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0299,"confidence_interval":0.946,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.54,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1688,"p95_latency_ms":2063,"p99_latency_ms":2213,"time_to_first_token_ms":307,"total_runtime_ms":1876,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0164,"confidence_interval":0.942,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0372,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1716,"p95_latency_ms":2098,"p99_latency_ms":2250,"time_to_first_token_ms":322,"total_runtime_ms":1907,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0291,"confidence_interval":0.942,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9635,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1023,"p95_latency_ms":1250,"p99_latency_ms":1341,"time_to_first_token_ms":281,"total_runtime_ms":1137,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0122,"confidence_interval":0.9367,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0725,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":969,"p95_latency_ms":1185,"p99_latency_ms":1271,"time_to_first_token_ms":290,"total_runtime_ms":1077,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9461,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1814,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":989,"p95_latency_ms":1209,"p99_latency_ms":1296,"time_to_first_token_ms":261,"total_runtime_ms":1099,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0125,"confidence_interval":0.9375,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":11.077,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1052,"p95_latency_ms":1285,"p99_latency_ms":1379,"time_to_first_token_ms":274,"total_runtime_ms":1169,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0277,"confidence_interval":0.9336,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6719,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1051,"p95_latency_ms":1285,"p99_latency_ms":1378,"time_to_first_token_ms":256,"total_runtime_ms":1168,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0162,"confidence_interval":0.9349,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8613,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1027,"p95_latency_ms":1255,"p99_latency_ms":1346,"time_to_first_token_ms":269,"total_runtime_ms":1141,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0187,"confidence_interval":0.9345,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5962,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":956,"p95_latency_ms":1169,"p99_latency_ms":1254,"time_to_first_token_ms":273,"total_runtime_ms":1063,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0195,"confidence_interval":0.9345,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9201,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":992,"p95_latency_ms":1212,"p99_latency_ms":1300,"time_to_first_token_ms":266,"total_runtime_ms":1102,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0151,"confidence_interval":0.94,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9531,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":994,"p95_latency_ms":1215,"p99_latency_ms":1303,"time_to_first_token_ms":275,"total_runtime_ms":1105,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0173,"confidence_interval":0.9431,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1533,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_strict_json|high|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":998,"p95_latency_ms":1220,"p99_latency_ms":1309,"time_to_first_token_ms":282,"total_runtime_ms":1109,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0122,"confidence_interval":0.9357,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0484,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1055,"p95_latency_ms":1289,"p99_latency_ms":1383,"time_to_first_token_ms":215,"total_runtime_ms":1172,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0121,"confidence_interval":0.9428,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2941,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1082,"p95_latency_ms":1323,"p99_latency_ms":1419,"time_to_first_token_ms":217,"total_runtime_ms":1203,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.945,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6108,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1067,"p95_latency_ms":1304,"p99_latency_ms":1399,"time_to_first_token_ms":211,"total_runtime_ms":1186,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0224,"confidence_interval":0.9334,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5003,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1121,"p95_latency_ms":1370,"p99_latency_ms":1470,"time_to_first_token_ms":217,"total_runtime_ms":1245,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0219,"confidence_interval":0.9329,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0827,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1148,"p95_latency_ms":1403,"p99_latency_ms":1505,"time_to_first_token_ms":189,"total_runtime_ms":1275,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0278,"confidence_interval":0.9402,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8836,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1114,"p95_latency_ms":1361,"p99_latency_ms":1460,"time_to_first_token_ms":211,"total_runtime_ms":1238,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0251,"confidence_interval":0.9472,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2118,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1134,"p95_latency_ms":1386,"p99_latency_ms":1487,"time_to_first_token_ms":217,"total_runtime_ms":1260,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0218,"confidence_interval":0.9471,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.92,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1087,"p95_latency_ms":1329,"p99_latency_ms":1426,"time_to_first_token_ms":190,"total_runtime_ms":1208,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0202,"confidence_interval":0.9387,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5597,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1118,"p95_latency_ms":1367,"p99_latency_ms":1466,"time_to_first_token_ms":202,"total_runtime_ms":1243,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0247,"confidence_interval":0.9469,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.672,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1100,"p95_latency_ms":1345,"p99_latency_ms":1443,"time_to_first_token_ms":207,"total_runtime_ms":1223,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0258,"confidence_interval":0.9348,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7359,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":741,"p95_latency_ms":905,"p99_latency_ms":971,"time_to_first_token_ms":176,"total_runtime_ms":823,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0235,"confidence_interval":0.9367,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8415,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":778,"p95_latency_ms":951,"p99_latency_ms":1020,"time_to_first_token_ms":174,"total_runtime_ms":864,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0167,"confidence_interval":0.9329,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6004,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":764,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":170,"total_runtime_ms":849,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0264,"confidence_interval":0.9466,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9645,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1001,"time_to_first_token_ms":157,"total_runtime_ms":849,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0255,"confidence_interval":0.9385,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3483,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":723,"p95_latency_ms":883,"p99_latency_ms":947,"time_to_first_token_ms":174,"total_runtime_ms":803,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0288,"confidence_interval":0.9373,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5589,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":754,"p95_latency_ms":922,"p99_latency_ms":989,"time_to_first_token_ms":173,"total_runtime_ms":838,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0248,"confidence_interval":0.9455,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5104,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":903,"time_to_first_token_ms":150,"total_runtime_ms":765,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0139,"confidence_interval":0.9318,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7888,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":696,"p95_latency_ms":850,"p99_latency_ms":912,"time_to_first_token_ms":161,"total_runtime_ms":773,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0275,"confidence_interval":0.9383,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8307,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":686,"p95_latency_ms":838,"p99_latency_ms":899,"time_to_first_token_ms":152,"total_runtime_ms":762,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0181,"confidence_interval":0.9425,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8365,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1003,"time_to_first_token_ms":164,"total_runtime_ms":850,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0107,"confidence_interval":0.9454,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0126,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1453,"p95_latency_ms":1776,"p99_latency_ms":1905,"time_to_first_token_ms":256,"total_runtime_ms":1614,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0154,"confidence_interval":0.9312,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8372,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1447,"p95_latency_ms":1768,"p99_latency_ms":1897,"time_to_first_token_ms":263,"total_runtime_ms":1607,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0281,"confidence_interval":0.9331,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.579,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1499,"p95_latency_ms":1832,"p99_latency_ms":1965,"time_to_first_token_ms":259,"total_runtime_ms":1665,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0122,"confidence_interval":0.9414,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.919,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1495,"p95_latency_ms":1828,"p99_latency_ms":1961,"time_to_first_token_ms":266,"total_runtime_ms":1662,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.024,"confidence_interval":0.9361,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3497,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1449,"p95_latency_ms":1771,"p99_latency_ms":1900,"time_to_first_token_ms":243,"total_runtime_ms":1610,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0162,"confidence_interval":0.9479,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9771,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1500,"p95_latency_ms":1833,"p99_latency_ms":1967,"time_to_first_token_ms":245,"total_runtime_ms":1667,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0289,"confidence_interval":0.945,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7154,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1504,"p95_latency_ms":1839,"p99_latency_ms":1972,"time_to_first_token_ms":251,"total_runtime_ms":1671,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.9452,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8431,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1513,"p95_latency_ms":1850,"p99_latency_ms":1984,"time_to_first_token_ms":234,"total_runtime_ms":1681,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0294,"confidence_interval":0.9375,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7894,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1460,"p95_latency_ms":1785,"p99_latency_ms":1914,"time_to_first_token_ms":247,"total_runtime_ms":1622,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0161,"confidence_interval":0.9398,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.56,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9064,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0936,"p50_latency_ms":1433,"p95_latency_ms":1752,"p99_latency_ms":1879,"time_to_first_token_ms":239,"total_runtime_ms":1592,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0102,"confidence_interval":0.9386,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0772,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":969,"p95_latency_ms":1184,"p99_latency_ms":1270,"time_to_first_token_ms":217,"total_runtime_ms":1076,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0242,"confidence_interval":0.9418,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1452,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":983,"p95_latency_ms":1201,"p99_latency_ms":1288,"time_to_first_token_ms":190,"total_runtime_ms":1092,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0263,"confidence_interval":0.9347,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.415,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":1008,"p95_latency_ms":1233,"p99_latency_ms":1322,"time_to_first_token_ms":215,"total_runtime_ms":1120,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0262,"confidence_interval":0.9354,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6708,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":956,"p95_latency_ms":1168,"p99_latency_ms":1253,"time_to_first_token_ms":195,"total_runtime_ms":1062,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0275,"confidence_interval":0.9437,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2193,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":973,"p95_latency_ms":1189,"p99_latency_ms":1276,"time_to_first_token_ms":218,"total_runtime_ms":1081,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0191,"confidence_interval":0.9393,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0902,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":1013,"p95_latency_ms":1238,"p99_latency_ms":1328,"time_to_first_token_ms":200,"total_runtime_ms":1126,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0219,"confidence_interval":0.9329,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6528,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":983,"p95_latency_ms":1202,"p99_latency_ms":1289,"time_to_first_token_ms":216,"total_runtime_ms":1092,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.014,"confidence_interval":0.9463,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2938,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":1008,"p95_latency_ms":1232,"p99_latency_ms":1322,"time_to_first_token_ms":197,"total_runtime_ms":1120,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0269,"confidence_interval":0.9436,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9193,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":971,"p95_latency_ms":1187,"p99_latency_ms":1273,"time_to_first_token_ms":213,"total_runtime_ms":1079,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0217,"confidence_interval":0.943,"validation_overhead_ms":106,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2694,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9064,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0936,"p50_latency_ms":996,"p95_latency_ms":1218,"p99_latency_ms":1306,"time_to_first_token_ms":221,"total_runtime_ms":1107,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0216,"confidence_interval":0.9422,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3917,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2887,"p95_latency_ms":3529,"p99_latency_ms":3786,"time_to_first_token_ms":434,"total_runtime_ms":3208,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0221,"confidence_interval":0.9397,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2121,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2866,"p95_latency_ms":3502,"p99_latency_ms":3757,"time_to_first_token_ms":421,"total_runtime_ms":3184,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.9362,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.344,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2896,"p95_latency_ms":3540,"p99_latency_ms":3797,"time_to_first_token_ms":435,"total_runtime_ms":3218,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0192,"confidence_interval":0.9334,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9069,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2887,"p95_latency_ms":3528,"p99_latency_ms":3785,"time_to_first_token_ms":427,"total_runtime_ms":3207,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0163,"confidence_interval":0.9439,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8584,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2850,"p95_latency_ms":3483,"p99_latency_ms":3736,"time_to_first_token_ms":412,"total_runtime_ms":3166,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0254,"confidence_interval":0.9309,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3638,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2832,"p95_latency_ms":3461,"p99_latency_ms":3712,"time_to_first_token_ms":416,"total_runtime_ms":3146,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0161,"confidence_interval":0.9442,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0743,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2857,"p95_latency_ms":3492,"p99_latency_ms":3746,"time_to_first_token_ms":424,"total_runtime_ms":3174,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0207,"confidence_interval":0.9434,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1543,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2877,"p95_latency_ms":3516,"p99_latency_ms":3772,"time_to_first_token_ms":429,"total_runtime_ms":3196,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0228,"confidence_interval":0.9481,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":9.155,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2889,"p95_latency_ms":3531,"p99_latency_ms":3788,"time_to_first_token_ms":416,"total_runtime_ms":3210,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0204,"confidence_interval":0.9384,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6301,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8848,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1152,"p50_latency_ms":2904,"p95_latency_ms":3549,"p99_latency_ms":3808,"time_to_first_token_ms":432,"total_runtime_ms":3227,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0169,"confidence_interval":0.948,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2994,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1905,"p95_latency_ms":2328,"p99_latency_ms":2498,"time_to_first_token_ms":372,"total_runtime_ms":2117,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0256,"confidence_interval":0.9394,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":8.809,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1883,"p95_latency_ms":2301,"p99_latency_ms":2468,"time_to_first_token_ms":384,"total_runtime_ms":2092,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0132,"confidence_interval":0.9428,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6465,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1841,"p95_latency_ms":2250,"p99_latency_ms":2414,"time_to_first_token_ms":394,"total_runtime_ms":2045,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0198,"confidence_interval":0.9396,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9636,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1863,"p95_latency_ms":2277,"p99_latency_ms":2443,"time_to_first_token_ms":374,"total_runtime_ms":2070,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0126,"confidence_interval":0.9376,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6224,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1919,"p95_latency_ms":2345,"p99_latency_ms":2516,"time_to_first_token_ms":382,"total_runtime_ms":2132,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0203,"confidence_interval":0.9468,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":9.3007,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1842,"p95_latency_ms":2251,"p99_latency_ms":2414,"time_to_first_token_ms":393,"total_runtime_ms":2046,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0273,"confidence_interval":0.9333,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2792,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1864,"p95_latency_ms":2279,"p99_latency_ms":2444,"time_to_first_token_ms":381,"total_runtime_ms":2072,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0266,"confidence_interval":0.9394,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6339,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1862,"p95_latency_ms":2276,"p99_latency_ms":2441,"time_to_first_token_ms":383,"total_runtime_ms":2069,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0218,"confidence_interval":0.9393,"validation_overhead_ms":125,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7591,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1837,"p95_latency_ms":2245,"p99_latency_ms":2409,"time_to_first_token_ms":371,"total_runtime_ms":2041,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0107,"confidence_interval":0.945,"validation_overhead_ms":131,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2884,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8848,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1152,"p50_latency_ms":1927,"p95_latency_ms":2356,"p99_latency_ms":2527,"time_to_first_token_ms":381,"total_runtime_ms":2142,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0263,"confidence_interval":0.9383,"validation_overhead_ms":119,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2092,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4264,"p95_latency_ms":5212,"p99_latency_ms":5591,"time_to_first_token_ms":585,"total_runtime_ms":4738,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0233,"confidence_interval":0.9353,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0834,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4206,"p95_latency_ms":5140,"p99_latency_ms":5514,"time_to_first_token_ms":581,"total_runtime_ms":4673,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0253,"confidence_interval":0.9477,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7941,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4283,"p95_latency_ms":5235,"p99_latency_ms":5616,"time_to_first_token_ms":600,"total_runtime_ms":4759,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0135,"confidence_interval":0.9388,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8475,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4256,"p95_latency_ms":5201,"p99_latency_ms":5580,"time_to_first_token_ms":587,"total_runtime_ms":4729,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0233,"confidence_interval":0.9404,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5026,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4297,"p95_latency_ms":5252,"p99_latency_ms":5634,"time_to_first_token_ms":601,"total_runtime_ms":4775,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0277,"confidence_interval":0.936,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0037,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4270,"p95_latency_ms":5219,"p99_latency_ms":5598,"time_to_first_token_ms":596,"total_runtime_ms":4744,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.022,"confidence_interval":0.9339,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5506,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4206,"p95_latency_ms":5140,"p99_latency_ms":5514,"time_to_first_token_ms":593,"total_runtime_ms":4673,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0198,"confidence_interval":0.9396,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7274,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4203,"p95_latency_ms":5137,"p99_latency_ms":5510,"time_to_first_token_ms":590,"total_runtime_ms":4670,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0285,"confidence_interval":0.9494,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5339,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4237,"p95_latency_ms":5178,"p99_latency_ms":5555,"time_to_first_token_ms":573,"total_runtime_ms":4708,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.022,"confidence_interval":0.9476,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9609,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.864,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.136,"p50_latency_ms":4206,"p95_latency_ms":5140,"p99_latency_ms":5514,"time_to_first_token_ms":589,"total_runtime_ms":4673,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0106,"confidence_interval":0.9412,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9409,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2758,"p95_latency_ms":3371,"p99_latency_ms":3616,"time_to_first_token_ms":545,"total_runtime_ms":3065,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0253,"confidence_interval":0.9475,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9975,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2765,"p95_latency_ms":3379,"p99_latency_ms":3625,"time_to_first_token_ms":548,"total_runtime_ms":3072,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.02,"confidence_interval":0.9378,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2158,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2724,"p95_latency_ms":3329,"p99_latency_ms":3571,"time_to_first_token_ms":534,"total_runtime_ms":3027,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0252,"confidence_interval":0.9386,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1334,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2749,"p95_latency_ms":3359,"p99_latency_ms":3604,"time_to_first_token_ms":551,"total_runtime_ms":3054,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0131,"confidence_interval":0.9336,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8682,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2712,"p95_latency_ms":3315,"p99_latency_ms":3556,"time_to_first_token_ms":548,"total_runtime_ms":3013,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0214,"confidence_interval":0.9381,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3022,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2720,"p95_latency_ms":3324,"p99_latency_ms":3566,"time_to_first_token_ms":531,"total_runtime_ms":3022,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0184,"confidence_interval":0.9301,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0334,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2754,"p95_latency_ms":3366,"p99_latency_ms":3611,"time_to_first_token_ms":533,"total_runtime_ms":3060,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0149,"confidence_interval":0.9435,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6189,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2804,"p95_latency_ms":3427,"p99_latency_ms":3676,"time_to_first_token_ms":543,"total_runtime_ms":3115,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0294,"confidence_interval":0.9363,"validation_overhead_ms":148,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9777,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2697,"p95_latency_ms":3297,"p99_latency_ms":3536,"time_to_first_token_ms":551,"total_runtime_ms":2997,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.013,"confidence_interval":0.94,"validation_overhead_ms":154,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0364,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|low|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.864,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.136,"p50_latency_ms":2698,"p95_latency_ms":3297,"p99_latency_ms":3537,"time_to_first_token_ms":535,"total_runtime_ms":2998,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0226,"confidence_interval":0.9371,"validation_overhead_ms":142,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1231,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":893,"p95_latency_ms":1092,"p99_latency_ms":1171,"time_to_first_token_ms":186,"total_runtime_ms":992,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0151,"confidence_interval":0.9387,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9889,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":842,"p95_latency_ms":1029,"p99_latency_ms":1104,"time_to_first_token_ms":179,"total_runtime_ms":935,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0164,"confidence_interval":0.9363,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.107,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":789,"p95_latency_ms":964,"p99_latency_ms":1034,"time_to_first_token_ms":175,"total_runtime_ms":876,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0172,"confidence_interval":0.9301,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8036,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":853,"p95_latency_ms":1043,"p99_latency_ms":1119,"time_to_first_token_ms":189,"total_runtime_ms":948,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0105,"confidence_interval":0.9386,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8248,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":872,"p95_latency_ms":1066,"p99_latency_ms":1143,"time_to_first_token_ms":181,"total_runtime_ms":969,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0284,"confidence_interval":0.9438,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7402,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":894,"p95_latency_ms":1092,"p99_latency_ms":1172,"time_to_first_token_ms":160,"total_runtime_ms":993,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0297,"confidence_interval":0.9301,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3643,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":892,"p95_latency_ms":1091,"p99_latency_ms":1170,"time_to_first_token_ms":163,"total_runtime_ms":992,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0128,"confidence_interval":0.9416,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.685,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":806,"p95_latency_ms":985,"p99_latency_ms":1057,"time_to_first_token_ms":177,"total_runtime_ms":896,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.018,"confidence_interval":0.9369,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2794,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":824,"p95_latency_ms":1008,"p99_latency_ms":1081,"time_to_first_token_ms":184,"total_runtime_ms":916,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0287,"confidence_interval":0.9321,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6085,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":816,"p95_latency_ms":997,"p99_latency_ms":1070,"time_to_first_token_ms":164,"total_runtime_ms":907,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0266,"confidence_interval":0.9452,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.632,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":539,"p95_latency_ms":658,"p99_latency_ms":706,"time_to_first_token_ms":121,"total_runtime_ms":599,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0134,"confidence_interval":0.9458,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2545,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":607,"p95_latency_ms":742,"p99_latency_ms":796,"time_to_first_token_ms":137,"total_runtime_ms":675,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.9428,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4006,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":535,"p95_latency_ms":654,"p99_latency_ms":701,"time_to_first_token_ms":145,"total_runtime_ms":594,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0274,"confidence_interval":0.9338,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5167,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":121,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0166,"confidence_interval":0.9327,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8667,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":507,"p95_latency_ms":620,"p99_latency_ms":665,"time_to_first_token_ms":112,"total_runtime_ms":564,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0135,"confidence_interval":0.9304,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2826,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":121,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.014,"confidence_interval":0.9375,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5537,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":565,"p95_latency_ms":691,"p99_latency_ms":741,"time_to_first_token_ms":127,"total_runtime_ms":628,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0224,"confidence_interval":0.941,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2851,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":538,"p95_latency_ms":658,"p99_latency_ms":705,"time_to_first_token_ms":125,"total_runtime_ms":598,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.029,"confidence_interval":0.9479,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1497,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":562,"p95_latency_ms":687,"p99_latency_ms":737,"time_to_first_token_ms":112,"total_runtime_ms":625,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0112,"confidence_interval":0.9311,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1679,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":558,"p95_latency_ms":683,"p99_latency_ms":732,"time_to_first_token_ms":121,"total_runtime_ms":620,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0259,"confidence_interval":0.949,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.751,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1125,"p95_latency_ms":1375,"p99_latency_ms":1475,"time_to_first_token_ms":192,"total_runtime_ms":1250,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0289,"confidence_interval":0.9477,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6531,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1099,"p95_latency_ms":1344,"p99_latency_ms":1441,"time_to_first_token_ms":218,"total_runtime_ms":1221,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.03,"confidence_interval":0.9433,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8744,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1169,"p95_latency_ms":1429,"p99_latency_ms":1533,"time_to_first_token_ms":193,"total_runtime_ms":1299,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0299,"confidence_interval":0.9439,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5733,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1110,"p95_latency_ms":1356,"p99_latency_ms":1455,"time_to_first_token_ms":208,"total_runtime_ms":1233,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0233,"confidence_interval":0.9492,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2958,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1162,"p95_latency_ms":1420,"p99_latency_ms":1524,"time_to_first_token_ms":203,"total_runtime_ms":1291,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9352,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0472,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1107,"p95_latency_ms":1353,"p99_latency_ms":1452,"time_to_first_token_ms":217,"total_runtime_ms":1230,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.026,"confidence_interval":0.9356,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7654,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1077,"p95_latency_ms":1316,"p99_latency_ms":1412,"time_to_first_token_ms":202,"total_runtime_ms":1197,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0232,"confidence_interval":0.9389,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1691,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1167,"p95_latency_ms":1427,"p99_latency_ms":1530,"time_to_first_token_ms":223,"total_runtime_ms":1297,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0277,"confidence_interval":0.947,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0372,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1080,"p95_latency_ms":1321,"p99_latency_ms":1417,"time_to_first_token_ms":220,"total_runtime_ms":1201,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.013,"confidence_interval":0.945,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3676,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9144,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0856,"p50_latency_ms":1169,"p95_latency_ms":1429,"p99_latency_ms":1533,"time_to_first_token_ms":197,"total_runtime_ms":1299,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0211,"confidence_interval":0.9432,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4928,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":726,"p95_latency_ms":887,"p99_latency_ms":952,"time_to_first_token_ms":152,"total_runtime_ms":807,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.02,"confidence_interval":0.9417,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1734,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":694,"p95_latency_ms":848,"p99_latency_ms":909,"time_to_first_token_ms":177,"total_runtime_ms":771,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0157,"confidence_interval":0.9375,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.882,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":785,"p95_latency_ms":959,"p99_latency_ms":1029,"time_to_first_token_ms":146,"total_runtime_ms":872,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0119,"confidence_interval":0.9399,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6621,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":780,"p95_latency_ms":954,"p99_latency_ms":1023,"time_to_first_token_ms":167,"total_runtime_ms":867,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0138,"confidence_interval":0.9387,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0472,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":684,"p95_latency_ms":835,"p99_latency_ms":896,"time_to_first_token_ms":153,"total_runtime_ms":759,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.018,"confidence_interval":0.9435,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.512,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":753,"p95_latency_ms":920,"p99_latency_ms":987,"time_to_first_token_ms":148,"total_runtime_ms":836,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0269,"confidence_interval":0.9402,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1533,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":768,"p95_latency_ms":938,"p99_latency_ms":1006,"time_to_first_token_ms":145,"total_runtime_ms":853,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0187,"confidence_interval":0.9437,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9242,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":707,"p95_latency_ms":864,"p99_latency_ms":927,"time_to_first_token_ms":148,"total_runtime_ms":786,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0233,"confidence_interval":0.9356,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5247,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":761,"p95_latency_ms":930,"p99_latency_ms":997,"time_to_first_token_ms":172,"total_runtime_ms":845,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.022,"confidence_interval":0.9368,"validation_overhead_ms":106,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2097,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9144,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0856,"p50_latency_ms":770,"p95_latency_ms":942,"p99_latency_ms":1010,"time_to_first_token_ms":146,"total_runtime_ms":856,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0206,"confidence_interval":0.9308,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6579,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2161,"p95_latency_ms":2641,"p99_latency_ms":2833,"time_to_first_token_ms":342,"total_runtime_ms":2401,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0201,"confidence_interval":0.9399,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2729,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2155,"p95_latency_ms":2634,"p99_latency_ms":2825,"time_to_first_token_ms":342,"total_runtime_ms":2394,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0249,"confidence_interval":0.9359,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8201,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2169,"p95_latency_ms":2651,"p99_latency_ms":2844,"time_to_first_token_ms":335,"total_runtime_ms":2410,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0133,"confidence_interval":0.935,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5098,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2204,"p95_latency_ms":2694,"p99_latency_ms":2890,"time_to_first_token_ms":339,"total_runtime_ms":2449,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0102,"confidence_interval":0.9429,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4377,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2217,"p95_latency_ms":2710,"p99_latency_ms":2907,"time_to_first_token_ms":340,"total_runtime_ms":2463,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0242,"confidence_interval":0.9484,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0293,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2221,"p95_latency_ms":2715,"p99_latency_ms":2912,"time_to_first_token_ms":323,"total_runtime_ms":2468,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0137,"confidence_interval":0.9332,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4414,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2125,"p95_latency_ms":2598,"p99_latency_ms":2787,"time_to_first_token_ms":337,"total_runtime_ms":2362,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0221,"confidence_interval":0.9402,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9437,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2164,"p95_latency_ms":2645,"p99_latency_ms":2838,"time_to_first_token_ms":349,"total_runtime_ms":2405,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0181,"confidence_interval":0.9473,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1224,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2185,"p95_latency_ms":2671,"p99_latency_ms":2865,"time_to_first_token_ms":332,"total_runtime_ms":2428,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0201,"confidence_interval":0.9326,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1717,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8928,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1072,"p50_latency_ms":2153,"p95_latency_ms":2632,"p99_latency_ms":2823,"time_to_first_token_ms":349,"total_runtime_ms":2393,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0135,"confidence_interval":0.9338,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.627,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1465,"p95_latency_ms":1791,"p99_latency_ms":1921,"time_to_first_token_ms":309,"total_runtime_ms":1628,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0113,"confidence_interval":0.9478,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2688,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1408,"p95_latency_ms":1721,"p99_latency_ms":1846,"time_to_first_token_ms":276,"total_runtime_ms":1564,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0151,"confidence_interval":0.9304,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9473,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1398,"p95_latency_ms":1709,"p99_latency_ms":1833,"time_to_first_token_ms":298,"total_runtime_ms":1554,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.01,"confidence_interval":0.934,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4397,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1370,"p95_latency_ms":1674,"p99_latency_ms":1796,"time_to_first_token_ms":292,"total_runtime_ms":1522,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0221,"confidence_interval":0.9419,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8234,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1375,"p95_latency_ms":1680,"p99_latency_ms":1802,"time_to_first_token_ms":289,"total_runtime_ms":1527,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0285,"confidence_interval":0.9313,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1133,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1438,"p95_latency_ms":1758,"p99_latency_ms":1886,"time_to_first_token_ms":296,"total_runtime_ms":1598,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.03,"confidence_interval":0.9351,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8576,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1381,"p95_latency_ms":1688,"p99_latency_ms":1810,"time_to_first_token_ms":278,"total_runtime_ms":1534,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.014,"confidence_interval":0.9363,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2179,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1433,"p95_latency_ms":1751,"p99_latency_ms":1879,"time_to_first_token_ms":306,"total_runtime_ms":1592,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0144,"confidence_interval":0.9477,"validation_overhead_ms":125,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0098,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1375,"p95_latency_ms":1680,"p99_latency_ms":1802,"time_to_first_token_ms":283,"total_runtime_ms":1527,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0218,"confidence_interval":0.9464,"validation_overhead_ms":131,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5748,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8928,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1072,"p50_latency_ms":1419,"p95_latency_ms":1734,"p99_latency_ms":1860,"time_to_first_token_ms":274,"total_runtime_ms":1576,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0265,"confidence_interval":0.9332,"validation_overhead_ms":119,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0399,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3154,"p95_latency_ms":3855,"p99_latency_ms":4135,"time_to_first_token_ms":471,"total_runtime_ms":3504,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0198,"confidence_interval":0.9449,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3217,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3179,"p95_latency_ms":3886,"p99_latency_ms":4168,"time_to_first_token_ms":471,"total_runtime_ms":3533,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.9322,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9129,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3227,"p95_latency_ms":3945,"p99_latency_ms":4231,"time_to_first_token_ms":464,"total_runtime_ms":3586,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0208,"confidence_interval":0.9323,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8382,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3216,"p95_latency_ms":3930,"p99_latency_ms":4216,"time_to_first_token_ms":470,"total_runtime_ms":3573,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.014,"confidence_interval":0.9438,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4944,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3202,"p95_latency_ms":3913,"p99_latency_ms":4198,"time_to_first_token_ms":451,"total_runtime_ms":3558,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.021,"confidence_interval":0.9325,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4823,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3164,"p95_latency_ms":3868,"p99_latency_ms":4149,"time_to_first_token_ms":476,"total_runtime_ms":3516,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0117,"confidence_interval":0.9399,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0443,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3218,"p95_latency_ms":3934,"p99_latency_ms":4220,"time_to_first_token_ms":453,"total_runtime_ms":3576,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0115,"confidence_interval":0.9497,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5852,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3242,"p95_latency_ms":3962,"p99_latency_ms":4250,"time_to_first_token_ms":444,"total_runtime_ms":3602,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0266,"confidence_interval":0.9382,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8904,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3193,"p95_latency_ms":3903,"p99_latency_ms":4187,"time_to_first_token_ms":444,"total_runtime_ms":3548,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0291,"confidence_interval":0.9371,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4344,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.872,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.128,"p50_latency_ms":3206,"p95_latency_ms":3919,"p99_latency_ms":4204,"time_to_first_token_ms":461,"total_runtime_ms":3563,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0177,"confidence_interval":0.9457,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2031,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2081,"p95_latency_ms":2543,"p99_latency_ms":2728,"time_to_first_token_ms":402,"total_runtime_ms":2312,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0221,"confidence_interval":0.9455,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4972,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2067,"p95_latency_ms":2527,"p99_latency_ms":2711,"time_to_first_token_ms":419,"total_runtime_ms":2297,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0226,"confidence_interval":0.9392,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8714,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2046,"p95_latency_ms":2501,"p99_latency_ms":2683,"time_to_first_token_ms":409,"total_runtime_ms":2274,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0204,"confidence_interval":0.9492,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":10.938,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2066,"p95_latency_ms":2526,"p99_latency_ms":2709,"time_to_first_token_ms":402,"total_runtime_ms":2296,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0197,"confidence_interval":0.9438,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.234,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2034,"p95_latency_ms":2486,"p99_latency_ms":2667,"time_to_first_token_ms":402,"total_runtime_ms":2260,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0215,"confidence_interval":0.9392,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8557,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2108,"p95_latency_ms":2576,"p99_latency_ms":2763,"time_to_first_token_ms":405,"total_runtime_ms":2342,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0225,"confidence_interval":0.9473,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4769,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2096,"p95_latency_ms":2561,"p99_latency_ms":2747,"time_to_first_token_ms":426,"total_runtime_ms":2328,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0137,"confidence_interval":0.9311,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1202,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2127,"p95_latency_ms":2600,"p99_latency_ms":2789,"time_to_first_token_ms":422,"total_runtime_ms":2364,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0222,"confidence_interval":0.937,"validation_overhead_ms":148,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3103,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2043,"p95_latency_ms":2497,"p99_latency_ms":2679,"time_to_first_token_ms":425,"total_runtime_ms":2270,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0224,"confidence_interval":0.9444,"validation_overhead_ms":154,"orchestration_overhead_ms":69,"event_dispatch_cost":11.079,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.872,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.128,"p50_latency_ms":2121,"p95_latency_ms":2593,"p99_latency_ms":2782,"time_to_first_token_ms":434,"total_runtime_ms":2357,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0189,"confidence_interval":0.93,"validation_overhead_ms":142,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8031,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":592,"p95_latency_ms":723,"p99_latency_ms":776,"time_to_first_token_ms":147,"total_runtime_ms":658,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0227,"confidence_interval":0.9352,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0908,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":614,"p95_latency_ms":750,"p99_latency_ms":805,"time_to_first_token_ms":144,"total_runtime_ms":682,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0141,"confidence_interval":0.9328,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9802,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":606,"p95_latency_ms":741,"p99_latency_ms":795,"time_to_first_token_ms":158,"total_runtime_ms":674,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0195,"confidence_interval":0.936,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8807,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":663,"p95_latency_ms":811,"p99_latency_ms":870,"time_to_first_token_ms":146,"total_runtime_ms":737,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0243,"confidence_interval":0.9394,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0862,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":610,"p95_latency_ms":746,"p99_latency_ms":800,"time_to_first_token_ms":146,"total_runtime_ms":678,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0284,"confidence_interval":0.948,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1164,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":646,"p95_latency_ms":790,"p99_latency_ms":847,"time_to_first_token_ms":137,"total_runtime_ms":718,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0292,"confidence_interval":0.9382,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3211,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":593,"p95_latency_ms":725,"p99_latency_ms":777,"time_to_first_token_ms":144,"total_runtime_ms":659,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0213,"confidence_interval":0.9381,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4509,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":610,"p95_latency_ms":745,"p99_latency_ms":799,"time_to_first_token_ms":153,"total_runtime_ms":677,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0107,"confidence_interval":0.9413,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3057,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":638,"p95_latency_ms":780,"p99_latency_ms":837,"time_to_first_token_ms":143,"total_runtime_ms":709,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.016,"confidence_interval":0.9382,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.573,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":689,"p95_latency_ms":842,"p99_latency_ms":904,"time_to_first_token_ms":150,"total_runtime_ms":766,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0297,"confidence_interval":0.9438,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6552,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":460,"p95_latency_ms":562,"p99_latency_ms":603,"time_to_first_token_ms":94,"total_runtime_ms":511,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0157,"confidence_interval":0.9324,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2456,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":403,"p95_latency_ms":493,"p99_latency_ms":529,"time_to_first_token_ms":114,"total_runtime_ms":448,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0236,"confidence_interval":0.9448,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2347,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":88,"total_runtime_ms":475,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0165,"confidence_interval":0.9409,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4197,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":406,"p95_latency_ms":496,"p99_latency_ms":532,"time_to_first_token_ms":121,"total_runtime_ms":451,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0202,"confidence_interval":0.9368,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4176,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":111,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.02,"confidence_interval":0.944,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.576,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":468,"p95_latency_ms":571,"p99_latency_ms":613,"time_to_first_token_ms":111,"total_runtime_ms":519,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0217,"confidence_interval":0.9381,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4734,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":449,"p95_latency_ms":548,"p99_latency_ms":588,"time_to_first_token_ms":97,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0106,"confidence_interval":0.9317,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6959,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":452,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":104,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0158,"confidence_interval":0.9355,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7257,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":395,"p95_latency_ms":483,"p99_latency_ms":518,"time_to_first_token_ms":105,"total_runtime_ms":439,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0216,"confidence_interval":0.9359,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1368,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":555,"time_to_first_token_ms":103,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0193,"confidence_interval":0.9434,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3705,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":896,"p95_latency_ms":1095,"p99_latency_ms":1174,"time_to_first_token_ms":191,"total_runtime_ms":995,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0298,"confidence_interval":0.9422,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0601,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":890,"p95_latency_ms":1087,"p99_latency_ms":1166,"time_to_first_token_ms":183,"total_runtime_ms":988,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0244,"confidence_interval":0.9436,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6201,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":822,"p95_latency_ms":1005,"p99_latency_ms":1078,"time_to_first_token_ms":160,"total_runtime_ms":913,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0137,"confidence_interval":0.9337,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0143,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":830,"p95_latency_ms":1014,"p99_latency_ms":1088,"time_to_first_token_ms":182,"total_runtime_ms":922,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.9413,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9126,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":819,"p95_latency_ms":1001,"p99_latency_ms":1074,"time_to_first_token_ms":168,"total_runtime_ms":910,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.019,"confidence_interval":0.9321,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7916,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":850,"p95_latency_ms":1039,"p99_latency_ms":1115,"time_to_first_token_ms":165,"total_runtime_ms":945,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0166,"confidence_interval":0.9437,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6656,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":802,"p95_latency_ms":980,"p99_latency_ms":1052,"time_to_first_token_ms":185,"total_runtime_ms":891,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0207,"confidence_interval":0.9447,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3389,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":851,"p95_latency_ms":1040,"p99_latency_ms":1116,"time_to_first_token_ms":172,"total_runtime_ms":945,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0163,"confidence_interval":0.9321,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6785,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":883,"p95_latency_ms":1079,"p99_latency_ms":1157,"time_to_first_token_ms":190,"total_runtime_ms":981,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0124,"confidence_interval":0.9317,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7194,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9204,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0796,"p50_latency_ms":835,"p95_latency_ms":1020,"p99_latency_ms":1095,"time_to_first_token_ms":162,"total_runtime_ms":928,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.941,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5723,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":531,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":114,"total_runtime_ms":591,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0244,"confidence_interval":0.9434,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0654,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":614,"p95_latency_ms":750,"p99_latency_ms":804,"time_to_first_token_ms":132,"total_runtime_ms":682,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0277,"confidence_interval":0.9416,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5842,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":608,"p95_latency_ms":743,"p99_latency_ms":797,"time_to_first_token_ms":117,"total_runtime_ms":676,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0186,"confidence_interval":0.9324,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6497,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":755,"time_to_first_token_ms":114,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0115,"confidence_interval":0.9398,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3508,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":547,"p95_latency_ms":668,"p99_latency_ms":717,"time_to_first_token_ms":118,"total_runtime_ms":608,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0184,"confidence_interval":0.9426,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7873,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":536,"p95_latency_ms":656,"p99_latency_ms":703,"time_to_first_token_ms":142,"total_runtime_ms":596,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9315,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8597,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":580,"p95_latency_ms":708,"p99_latency_ms":760,"time_to_first_token_ms":131,"total_runtime_ms":644,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.9499,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2306,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":592,"p95_latency_ms":724,"p99_latency_ms":776,"time_to_first_token_ms":119,"total_runtime_ms":658,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0298,"confidence_interval":0.9334,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2295,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":144,"total_runtime_ms":591,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0118,"confidence_interval":0.9311,"validation_overhead_ms":106,"orchestration_overhead_ms":66,"event_dispatch_cost":6.182,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9204,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0796,"p50_latency_ms":562,"p95_latency_ms":687,"p99_latency_ms":737,"time_to_first_token_ms":137,"total_runtime_ms":624,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0216,"confidence_interval":0.9345,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5697,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1599,"p95_latency_ms":1955,"p99_latency_ms":2097,"time_to_first_token_ms":272,"total_runtime_ms":1777,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0108,"confidence_interval":0.9468,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2797,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1615,"p95_latency_ms":1974,"p99_latency_ms":2118,"time_to_first_token_ms":275,"total_runtime_ms":1795,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0268,"confidence_interval":0.9483,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0471,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1596,"p95_latency_ms":1950,"p99_latency_ms":2092,"time_to_first_token_ms":258,"total_runtime_ms":1773,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0222,"confidence_interval":0.9339,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6558,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1604,"p95_latency_ms":1961,"p99_latency_ms":2103,"time_to_first_token_ms":256,"total_runtime_ms":1783,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0169,"confidence_interval":0.9339,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4933,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1641,"p95_latency_ms":2006,"p99_latency_ms":2152,"time_to_first_token_ms":287,"total_runtime_ms":1824,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0188,"confidence_interval":0.9353,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4284,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1675,"p95_latency_ms":2047,"p99_latency_ms":2196,"time_to_first_token_ms":282,"total_runtime_ms":1861,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0233,"confidence_interval":0.943,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7886,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1670,"p95_latency_ms":2041,"p99_latency_ms":2189,"time_to_first_token_ms":272,"total_runtime_ms":1855,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0223,"confidence_interval":0.9326,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6122,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1671,"p95_latency_ms":2042,"p99_latency_ms":2191,"time_to_first_token_ms":268,"total_runtime_ms":1857,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0181,"confidence_interval":0.9322,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3669,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1673,"p95_latency_ms":2045,"p99_latency_ms":2194,"time_to_first_token_ms":279,"total_runtime_ms":1859,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0186,"confidence_interval":0.9496,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9503,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.8988,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.1012,"p50_latency_ms":1678,"p95_latency_ms":2051,"p99_latency_ms":2201,"time_to_first_token_ms":288,"total_runtime_ms":1865,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0178,"confidence_interval":0.944,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9592,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1080,"p95_latency_ms":1320,"p99_latency_ms":1416,"time_to_first_token_ms":234,"total_runtime_ms":1200,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0128,"confidence_interval":0.9412,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4377,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1046,"p95_latency_ms":1279,"p99_latency_ms":1372,"time_to_first_token_ms":211,"total_runtime_ms":1163,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0187,"confidence_interval":0.9431,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4055,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1068,"p95_latency_ms":1305,"p99_latency_ms":1400,"time_to_first_token_ms":219,"total_runtime_ms":1186,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0257,"confidence_interval":0.9474,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6004,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1047,"p95_latency_ms":1279,"p99_latency_ms":1372,"time_to_first_token_ms":244,"total_runtime_ms":1163,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0258,"confidence_interval":0.9426,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1628,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1063,"p95_latency_ms":1299,"p99_latency_ms":1394,"time_to_first_token_ms":233,"total_runtime_ms":1181,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0179,"confidence_interval":0.9391,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7749,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1034,"p95_latency_ms":1264,"p99_latency_ms":1355,"time_to_first_token_ms":223,"total_runtime_ms":1149,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0225,"confidence_interval":0.9425,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3528,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1080,"p95_latency_ms":1320,"p99_latency_ms":1416,"time_to_first_token_ms":243,"total_runtime_ms":1200,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0119,"confidence_interval":0.9492,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2347,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1065,"p95_latency_ms":1302,"p99_latency_ms":1397,"time_to_first_token_ms":229,"total_runtime_ms":1184,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9479,"validation_overhead_ms":125,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1608,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1131,"p95_latency_ms":1383,"p99_latency_ms":1483,"time_to_first_token_ms":239,"total_runtime_ms":1257,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.9396,"validation_overhead_ms":131,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7882,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.8988,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.1012,"p50_latency_ms":1083,"p95_latency_ms":1324,"p99_latency_ms":1420,"time_to_first_token_ms":236,"total_runtime_ms":1204,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0151,"confidence_interval":0.9414,"validation_overhead_ms":119,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4035,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2448,"p95_latency_ms":2992,"p99_latency_ms":3210,"time_to_first_token_ms":375,"total_runtime_ms":2720,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0286,"confidence_interval":0.9488,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1016,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2391,"p95_latency_ms":2923,"p99_latency_ms":3135,"time_to_first_token_ms":369,"total_runtime_ms":2657,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0268,"confidence_interval":0.9441,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5792,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2426,"p95_latency_ms":2965,"p99_latency_ms":3181,"time_to_first_token_ms":374,"total_runtime_ms":2696,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0238,"confidence_interval":0.947,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5252,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2392,"p95_latency_ms":2924,"p99_latency_ms":3137,"time_to_first_token_ms":370,"total_runtime_ms":2658,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0215,"confidence_interval":0.9375,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0466,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2450,"p95_latency_ms":2995,"p99_latency_ms":3212,"time_to_first_token_ms":363,"total_runtime_ms":2722,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.02,"confidence_interval":0.9414,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0281,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2447,"p95_latency_ms":2990,"p99_latency_ms":3208,"time_to_first_token_ms":364,"total_runtime_ms":2719,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.011,"confidence_interval":0.9389,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2657,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2417,"p95_latency_ms":2955,"p99_latency_ms":3169,"time_to_first_token_ms":349,"total_runtime_ms":2686,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0186,"confidence_interval":0.9335,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5114,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2413,"p95_latency_ms":2950,"p99_latency_ms":3164,"time_to_first_token_ms":352,"total_runtime_ms":2681,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0278,"confidence_interval":0.9354,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2551,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2424,"p95_latency_ms":2963,"p99_latency_ms":3179,"time_to_first_token_ms":376,"total_runtime_ms":2694,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0279,"confidence_interval":0.9361,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6509,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.878,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.122,"p50_latency_ms":2438,"p95_latency_ms":2980,"p99_latency_ms":3196,"time_to_first_token_ms":355,"total_runtime_ms":2709,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0191,"confidence_interval":0.9456,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9854,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1531,"p95_latency_ms":1871,"p99_latency_ms":2007,"time_to_first_token_ms":314,"total_runtime_ms":1701,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0179,"confidence_interval":0.9496,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5479,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1595,"p95_latency_ms":1950,"p99_latency_ms":2092,"time_to_first_token_ms":319,"total_runtime_ms":1773,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0147,"confidence_interval":0.9446,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2717,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1591,"p95_latency_ms":1945,"p99_latency_ms":2086,"time_to_first_token_ms":304,"total_runtime_ms":1768,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0158,"confidence_interval":0.9499,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5634,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1623,"p95_latency_ms":1984,"p99_latency_ms":2128,"time_to_first_token_ms":314,"total_runtime_ms":1803,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0271,"confidence_interval":0.9358,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.466,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1591,"p95_latency_ms":1945,"p99_latency_ms":2086,"time_to_first_token_ms":317,"total_runtime_ms":1768,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0151,"confidence_interval":0.9398,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6682,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1542,"p95_latency_ms":1885,"p99_latency_ms":2022,"time_to_first_token_ms":321,"total_runtime_ms":1713,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0225,"confidence_interval":0.9354,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2829,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1547,"p95_latency_ms":1891,"p99_latency_ms":2028,"time_to_first_token_ms":306,"total_runtime_ms":1719,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0142,"confidence_interval":0.9403,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7951,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1528,"p95_latency_ms":1867,"p99_latency_ms":2003,"time_to_first_token_ms":331,"total_runtime_ms":1698,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0174,"confidence_interval":0.9325,"validation_overhead_ms":148,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4639,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1612,"p95_latency_ms":1971,"p99_latency_ms":2114,"time_to_first_token_ms":305,"total_runtime_ms":1792,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9329,"validation_overhead_ms":154,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5592,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.878,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.122,"p50_latency_ms":1553,"p95_latency_ms":1898,"p99_latency_ms":2036,"time_to_first_token_ms":323,"total_runtime_ms":1726,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0224,"confidence_interval":0.9378,"validation_overhead_ms":142,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4115,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":546,"p95_latency_ms":667,"p99_latency_ms":716,"time_to_first_token_ms":144,"total_runtime_ms":607,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9377,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2854,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":768,"time_to_first_token_ms":153,"total_runtime_ms":650,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9409,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4074,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":499,"p95_latency_ms":610,"p99_latency_ms":655,"time_to_first_token_ms":154,"total_runtime_ms":555,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0209,"confidence_interval":0.9461,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9254,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":152,"total_runtime_ms":664,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0238,"confidence_interval":0.9369,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8123,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":671,"time_to_first_token_ms":148,"total_runtime_ms":569,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0264,"confidence_interval":0.9419,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7732,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":592,"p95_latency_ms":724,"p99_latency_ms":776,"time_to_first_token_ms":141,"total_runtime_ms":658,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0187,"confidence_interval":0.9313,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9276,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":524,"p95_latency_ms":640,"p99_latency_ms":687,"time_to_first_token_ms":144,"total_runtime_ms":582,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0208,"confidence_interval":0.9455,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1994,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":590,"p95_latency_ms":721,"p99_latency_ms":773,"time_to_first_token_ms":124,"total_runtime_ms":655,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0185,"confidence_interval":0.9323,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4533,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":583,"p95_latency_ms":712,"p99_latency_ms":764,"time_to_first_token_ms":131,"total_runtime_ms":647,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0118,"confidence_interval":0.9318,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4195,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":543,"p95_latency_ms":663,"p99_latency_ms":712,"time_to_first_token_ms":122,"total_runtime_ms":603,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0156,"confidence_interval":0.9337,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0077,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":409,"p95_latency_ms":500,"p99_latency_ms":536,"time_to_first_token_ms":99,"total_runtime_ms":455,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0174,"confidence_interval":0.9349,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3997,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":496,"time_to_first_token_ms":85,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0112,"confidence_interval":0.943,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3315,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":483,"time_to_first_token_ms":91,"total_runtime_ms":409,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0237,"confidence_interval":0.9355,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9572,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":367,"p95_latency_ms":449,"p99_latency_ms":481,"time_to_first_token_ms":87,"total_runtime_ms":408,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0256,"confidence_interval":0.9395,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0576,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":383,"p95_latency_ms":468,"p99_latency_ms":502,"time_to_first_token_ms":92,"total_runtime_ms":425,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.018,"confidence_interval":0.9311,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4928,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":361,"p95_latency_ms":441,"p99_latency_ms":473,"time_to_first_token_ms":94,"total_runtime_ms":401,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0226,"confidence_interval":0.9324,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4347,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":411,"p95_latency_ms":503,"p99_latency_ms":539,"time_to_first_token_ms":88,"total_runtime_ms":457,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0153,"confidence_interval":0.9335,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2584,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":354,"p95_latency_ms":433,"p99_latency_ms":465,"time_to_first_token_ms":94,"total_runtime_ms":394,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0236,"confidence_interval":0.9343,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.731,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":338,"p95_latency_ms":413,"p99_latency_ms":443,"time_to_first_token_ms":104,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0265,"confidence_interval":0.9351,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9032,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":351,"p95_latency_ms":429,"p99_latency_ms":460,"time_to_first_token_ms":110,"total_runtime_ms":390,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.22,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0275,"confidence_interval":0.9472,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8442,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":714,"p95_latency_ms":872,"p99_latency_ms":936,"time_to_first_token_ms":162,"total_runtime_ms":793,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0192,"confidence_interval":0.9311,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4436,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":759,"p95_latency_ms":928,"p99_latency_ms":995,"time_to_first_token_ms":163,"total_runtime_ms":843,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0258,"confidence_interval":0.9381,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6762,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":176,"total_runtime_ms":849,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0227,"confidence_interval":0.936,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0296,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":679,"p95_latency_ms":830,"p99_latency_ms":890,"time_to_first_token_ms":142,"total_runtime_ms":754,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.024,"confidence_interval":0.9341,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0362,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":684,"p95_latency_ms":836,"p99_latency_ms":897,"time_to_first_token_ms":150,"total_runtime_ms":760,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0109,"confidence_interval":0.9379,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7832,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":737,"p95_latency_ms":900,"p99_latency_ms":966,"time_to_first_token_ms":153,"total_runtime_ms":819,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0174,"confidence_interval":0.9308,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":7.135,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":711,"p95_latency_ms":869,"p99_latency_ms":932,"time_to_first_token_ms":176,"total_runtime_ms":790,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0108,"confidence_interval":0.9405,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3953,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":738,"p95_latency_ms":901,"p99_latency_ms":967,"time_to_first_token_ms":168,"total_runtime_ms":819,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0196,"confidence_interval":0.9371,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1238,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":703,"p95_latency_ms":859,"p99_latency_ms":922,"time_to_first_token_ms":155,"total_runtime_ms":781,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0206,"confidence_interval":0.9399,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9229,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9234,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0766,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1003,"time_to_first_token_ms":168,"total_runtime_ms":850,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0107,"confidence_interval":0.9341,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0664,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":524,"p95_latency_ms":641,"p99_latency_ms":688,"time_to_first_token_ms":119,"total_runtime_ms":583,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0135,"confidence_interval":0.9309,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.773,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":486,"p95_latency_ms":594,"p99_latency_ms":637,"time_to_first_token_ms":119,"total_runtime_ms":540,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.011,"confidence_interval":0.9352,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2334,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":522,"p95_latency_ms":638,"p99_latency_ms":685,"time_to_first_token_ms":131,"total_runtime_ms":580,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0195,"confidence_interval":0.9386,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3326,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":465,"p95_latency_ms":568,"p99_latency_ms":609,"time_to_first_token_ms":104,"total_runtime_ms":516,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0145,"confidence_interval":0.934,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8901,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":470,"p95_latency_ms":574,"p99_latency_ms":616,"time_to_first_token_ms":118,"total_runtime_ms":522,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0269,"confidence_interval":0.9303,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.602,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":511,"p95_latency_ms":624,"p99_latency_ms":669,"time_to_first_token_ms":126,"total_runtime_ms":567,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0284,"confidence_interval":0.9325,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8435,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":508,"p95_latency_ms":621,"p99_latency_ms":666,"time_to_first_token_ms":110,"total_runtime_ms":565,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0251,"confidence_interval":0.9451,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2943,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":444,"p95_latency_ms":542,"p99_latency_ms":582,"time_to_first_token_ms":104,"total_runtime_ms":493,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0131,"confidence_interval":0.9393,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9615,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":452,"p95_latency_ms":552,"p99_latency_ms":592,"time_to_first_token_ms":120,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0144,"confidence_interval":0.9463,"validation_overhead_ms":106,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3391,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9234,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0766,"p50_latency_ms":495,"p95_latency_ms":605,"p99_latency_ms":649,"time_to_first_token_ms":100,"total_runtime_ms":550,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.1,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0266,"confidence_interval":0.9352,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0394,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1379,"p95_latency_ms":1685,"p99_latency_ms":1808,"time_to_first_token_ms":241,"total_runtime_ms":1532,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0189,"confidence_interval":0.9383,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":9.192,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1345,"p95_latency_ms":1644,"p99_latency_ms":1764,"time_to_first_token_ms":235,"total_runtime_ms":1495,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0106,"confidence_interval":0.9416,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6008,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1388,"p95_latency_ms":1696,"p99_latency_ms":1819,"time_to_first_token_ms":240,"total_runtime_ms":1542,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0236,"confidence_interval":0.9481,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4929,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1372,"p95_latency_ms":1677,"p99_latency_ms":1799,"time_to_first_token_ms":223,"total_runtime_ms":1524,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0139,"confidence_interval":0.9399,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6322,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1356,"p95_latency_ms":1658,"p99_latency_ms":1778,"time_to_first_token_ms":224,"total_runtime_ms":1507,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0178,"confidence_interval":0.9475,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8655,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1338,"p95_latency_ms":1636,"p99_latency_ms":1755,"time_to_first_token_ms":247,"total_runtime_ms":1487,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0277,"confidence_interval":0.9407,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0704,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1367,"p95_latency_ms":1670,"p99_latency_ms":1792,"time_to_first_token_ms":230,"total_runtime_ms":1518,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9349,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4312,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1414,"p95_latency_ms":1728,"p99_latency_ms":1854,"time_to_first_token_ms":240,"total_runtime_ms":1571,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0119,"confidence_interval":0.9352,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0206,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1356,"p95_latency_ms":1657,"p99_latency_ms":1777,"time_to_first_token_ms":245,"total_runtime_ms":1506,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0258,"confidence_interval":0.9372,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7366,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9018,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0982,"p50_latency_ms":1418,"p95_latency_ms":1733,"p99_latency_ms":1860,"time_to_first_token_ms":250,"total_runtime_ms":1576,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0119,"confidence_interval":0.9417,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9033,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":875,"p95_latency_ms":1069,"p99_latency_ms":1147,"time_to_first_token_ms":207,"total_runtime_ms":972,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0236,"confidence_interval":0.941,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4463,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":856,"p95_latency_ms":1046,"p99_latency_ms":1123,"time_to_first_token_ms":195,"total_runtime_ms":951,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.9316,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4662,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":947,"p95_latency_ms":1157,"p99_latency_ms":1242,"time_to_first_token_ms":202,"total_runtime_ms":1052,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0255,"confidence_interval":0.9374,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4703,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":898,"p95_latency_ms":1098,"p99_latency_ms":1177,"time_to_first_token_ms":203,"total_runtime_ms":998,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.021,"confidence_interval":0.9413,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.329,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":882,"p95_latency_ms":1078,"p99_latency_ms":1156,"time_to_first_token_ms":199,"total_runtime_ms":980,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0185,"confidence_interval":0.9372,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3786,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":922,"p95_latency_ms":1127,"p99_latency_ms":1209,"time_to_first_token_ms":196,"total_runtime_ms":1025,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0215,"confidence_interval":0.9308,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6778,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":900,"p95_latency_ms":1100,"p99_latency_ms":1180,"time_to_first_token_ms":187,"total_runtime_ms":1000,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0116,"confidence_interval":0.9467,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2541,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":955,"p95_latency_ms":1167,"p99_latency_ms":1252,"time_to_first_token_ms":209,"total_runtime_ms":1061,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0254,"confidence_interval":0.9307,"validation_overhead_ms":125,"orchestration_overhead_ms":65,"event_dispatch_cost":9.288,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":922,"p95_latency_ms":1127,"p99_latency_ms":1209,"time_to_first_token_ms":211,"total_runtime_ms":1024,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0237,"confidence_interval":0.9499,"validation_overhead_ms":131,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2538,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:7b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9018,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0982,"p50_latency_ms":909,"p95_latency_ms":1111,"p99_latency_ms":1191,"time_to_first_token_ms":190,"total_runtime_ms":1010,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0119,"confidence_interval":0.9487,"validation_overhead_ms":119,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9348,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":2006,"p95_latency_ms":2452,"p99_latency_ms":2631,"time_to_first_token_ms":322,"total_runtime_ms":2229,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0187,"confidence_interval":0.9361,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5007,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":2014,"p95_latency_ms":2462,"p99_latency_ms":2641,"time_to_first_token_ms":305,"total_runtime_ms":2238,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0287,"confidence_interval":0.9394,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0438,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":2032,"p95_latency_ms":2483,"p99_latency_ms":2664,"time_to_first_token_ms":327,"total_runtime_ms":2258,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0292,"confidence_interval":0.9325,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":11.003,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":2053,"p95_latency_ms":2510,"p99_latency_ms":2692,"time_to_first_token_ms":321,"total_runtime_ms":2282,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0125,"confidence_interval":0.9449,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8181,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1984,"p95_latency_ms":2425,"p99_latency_ms":2602,"time_to_first_token_ms":320,"total_runtime_ms":2205,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.029,"confidence_interval":0.9442,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2737,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":2015,"p95_latency_ms":2463,"p99_latency_ms":2642,"time_to_first_token_ms":324,"total_runtime_ms":2239,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0194,"confidence_interval":0.9381,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":10.879,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1970,"p95_latency_ms":2407,"p99_latency_ms":2582,"time_to_first_token_ms":333,"total_runtime_ms":2188,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0124,"confidence_interval":0.9406,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5157,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1991,"p95_latency_ms":2433,"p99_latency_ms":2610,"time_to_first_token_ms":307,"total_runtime_ms":2212,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.018,"confidence_interval":0.9452,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5901,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1971,"p95_latency_ms":2409,"p99_latency_ms":2584,"time_to_first_token_ms":316,"total_runtime_ms":2190,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0118,"confidence_interval":0.934,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9339,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.881,"parse_success_rate":0.892,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.119,"p50_latency_ms":1989,"p95_latency_ms":2431,"p99_latency_ms":2608,"time_to_first_token_ms":309,"total_runtime_ms":2210,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0253,"confidence_interval":0.9391,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1781,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1293,"p95_latency_ms":1580,"p99_latency_ms":1695,"time_to_first_token_ms":277,"total_runtime_ms":1436,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0174,"confidence_interval":0.9358,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5032,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1369,"p95_latency_ms":1673,"p99_latency_ms":1794,"time_to_first_token_ms":289,"total_runtime_ms":1521,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0103,"confidence_interval":0.9461,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6451,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1352,"p95_latency_ms":1653,"p99_latency_ms":1773,"time_to_first_token_ms":257,"total_runtime_ms":1502,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0293,"confidence_interval":0.9409,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3133,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1338,"p95_latency_ms":1636,"p99_latency_ms":1755,"time_to_first_token_ms":258,"total_runtime_ms":1487,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0222,"confidence_interval":0.9423,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":11.213,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1317,"p95_latency_ms":1610,"p99_latency_ms":1727,"time_to_first_token_ms":287,"total_runtime_ms":1464,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0187,"confidence_interval":0.9305,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8887,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1268,"p95_latency_ms":1550,"p99_latency_ms":1663,"time_to_first_token_ms":259,"total_runtime_ms":1409,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0189,"confidence_interval":0.9351,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4011,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1280,"p95_latency_ms":1565,"p99_latency_ms":1679,"time_to_first_token_ms":287,"total_runtime_ms":1422,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0249,"confidence_interval":0.9465,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6895,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1301,"p95_latency_ms":1590,"p99_latency_ms":1706,"time_to_first_token_ms":276,"total_runtime_ms":1446,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0123,"confidence_interval":0.9406,"validation_overhead_ms":148,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8404,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1316,"p95_latency_ms":1608,"p99_latency_ms":1725,"time_to_first_token_ms":275,"total_runtime_ms":1462,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0146,"confidence_interval":0.9347,"validation_overhead_ms":154,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1725,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"extract_nested_json|high|qwen2:14b|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.881,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.119,"p50_latency_ms":1270,"p95_latency_ms":1552,"p99_latency_ms":1665,"time_to_first_token_ms":274,"total_runtime_ms":1411,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0193,"confidence_interval":0.9421,"validation_overhead_ms":142,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4309,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":763,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":188,"total_runtime_ms":848,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0187,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9452,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":821,"p95_latency_ms":1003,"p99_latency_ms":1076,"time_to_first_token_ms":201,"total_runtime_ms":912,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0209,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.205,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":788,"p95_latency_ms":963,"p99_latency_ms":1033,"time_to_first_token_ms":200,"total_runtime_ms":875,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.025,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2956,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":846,"p95_latency_ms":1034,"p99_latency_ms":1110,"time_to_first_token_ms":204,"total_runtime_ms":940,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0201,"confidence_interval":0.9451,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.81,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":788,"p95_latency_ms":964,"p99_latency_ms":1034,"time_to_first_token_ms":219,"total_runtime_ms":876,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0144,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.253,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":776,"p95_latency_ms":948,"p99_latency_ms":1017,"time_to_first_token_ms":188,"total_runtime_ms":862,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0169,"confidence_interval":0.9331,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3197,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":831,"p95_latency_ms":1015,"p99_latency_ms":1089,"time_to_first_token_ms":221,"total_runtime_ms":923,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0177,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0622,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":767,"p95_latency_ms":938,"p99_latency_ms":1006,"time_to_first_token_ms":204,"total_runtime_ms":853,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0276,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.979,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":817,"p95_latency_ms":998,"p99_latency_ms":1071,"time_to_first_token_ms":192,"total_runtime_ms":907,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0162,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0244,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":815,"p95_latency_ms":996,"p99_latency_ms":1068,"time_to_first_token_ms":222,"total_runtime_ms":905,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0154,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9322,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":394,"p95_latency_ms":482,"p99_latency_ms":517,"time_to_first_token_ms":163,"total_runtime_ms":438,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.026,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4496,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":443,"p95_latency_ms":541,"p99_latency_ms":581,"time_to_first_token_ms":152,"total_runtime_ms":492,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0174,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7113,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":473,"p95_latency_ms":578,"p99_latency_ms":621,"time_to_first_token_ms":151,"total_runtime_ms":526,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0235,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7381,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":464,"p95_latency_ms":567,"p99_latency_ms":608,"time_to_first_token_ms":156,"total_runtime_ms":515,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.017,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2421,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":448,"p95_latency_ms":547,"p99_latency_ms":587,"time_to_first_token_ms":151,"total_runtime_ms":498,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0112,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.44,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":174,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0246,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9759,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":446,"p95_latency_ms":545,"p99_latency_ms":584,"time_to_first_token_ms":151,"total_runtime_ms":495,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0239,"confidence_interval":0.9408,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1105,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":479,"p95_latency_ms":585,"p99_latency_ms":628,"time_to_first_token_ms":144,"total_runtime_ms":532,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0152,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3803,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":456,"p95_latency_ms":558,"p99_latency_ms":598,"time_to_first_token_ms":169,"total_runtime_ms":507,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0135,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4317,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":548,"time_to_first_token_ms":160,"total_runtime_ms":465,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0254,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1122,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1096,"p95_latency_ms":1339,"p99_latency_ms":1436,"time_to_first_token_ms":260,"total_runtime_ms":1217,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.317,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1094,"p95_latency_ms":1337,"p99_latency_ms":1434,"time_to_first_token_ms":253,"total_runtime_ms":1215,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0287,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.895,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1344,"time_to_first_token_ms":238,"total_runtime_ms":1139,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0238,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2609,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1116,"p95_latency_ms":1364,"p99_latency_ms":1463,"time_to_first_token_ms":247,"total_runtime_ms":1240,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.023,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2833,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1106,"p95_latency_ms":1352,"p99_latency_ms":1450,"time_to_first_token_ms":235,"total_runtime_ms":1229,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.03,"confidence_interval":0.9411,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0437,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1034,"p95_latency_ms":1264,"p99_latency_ms":1356,"time_to_first_token_ms":248,"total_runtime_ms":1149,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0163,"confidence_interval":0.9342,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0588,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1086,"p95_latency_ms":1328,"p99_latency_ms":1424,"time_to_first_token_ms":249,"total_runtime_ms":1207,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0147,"confidence_interval":0.9495,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1149,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1079,"p95_latency_ms":1318,"p99_latency_ms":1414,"time_to_first_token_ms":243,"total_runtime_ms":1198,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0114,"confidence_interval":0.9393,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9261,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1102,"p95_latency_ms":1347,"p99_latency_ms":1445,"time_to_first_token_ms":242,"total_runtime_ms":1224,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0173,"confidence_interval":0.9455,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9535,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1040,"p95_latency_ms":1271,"p99_latency_ms":1363,"time_to_first_token_ms":264,"total_runtime_ms":1156,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0291,"confidence_interval":0.9445,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.702,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":539,"p95_latency_ms":659,"p99_latency_ms":707,"time_to_first_token_ms":192,"total_runtime_ms":599,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0104,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9257,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":610,"p95_latency_ms":746,"p99_latency_ms":800,"time_to_first_token_ms":215,"total_runtime_ms":678,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0288,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9825,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":581,"p95_latency_ms":710,"p99_latency_ms":761,"time_to_first_token_ms":201,"total_runtime_ms":645,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0204,"confidence_interval":0.9344,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1836,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":545,"p95_latency_ms":666,"p99_latency_ms":714,"time_to_first_token_ms":219,"total_runtime_ms":605,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0234,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0381,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":519,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":219,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0196,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5725,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":574,"p95_latency_ms":702,"p99_latency_ms":753,"time_to_first_token_ms":189,"total_runtime_ms":638,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0145,"confidence_interval":0.9381,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6337,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":568,"p95_latency_ms":694,"p99_latency_ms":744,"time_to_first_token_ms":191,"total_runtime_ms":631,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0268,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1594,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":578,"p95_latency_ms":707,"p99_latency_ms":758,"time_to_first_token_ms":190,"total_runtime_ms":642,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0108,"confidence_interval":0.9485,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7223,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":586,"p95_latency_ms":716,"p99_latency_ms":768,"time_to_first_token_ms":205,"total_runtime_ms":651,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0202,"confidence_interval":0.9308,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7488,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":564,"p95_latency_ms":689,"p99_latency_ms":739,"time_to_first_token_ms":218,"total_runtime_ms":626,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0137,"confidence_interval":0.9497,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5005,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2052,"p95_latency_ms":2509,"p99_latency_ms":2691,"time_to_first_token_ms":429,"total_runtime_ms":2281,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0275,"confidence_interval":0.9483,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0985,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2047,"p95_latency_ms":2502,"p99_latency_ms":2683,"time_to_first_token_ms":435,"total_runtime_ms":2274,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0281,"confidence_interval":0.9341,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1877,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2106,"p95_latency_ms":2574,"p99_latency_ms":2761,"time_to_first_token_ms":429,"total_runtime_ms":2340,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0151,"confidence_interval":0.9469,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4223,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2126,"p95_latency_ms":2599,"p99_latency_ms":2788,"time_to_first_token_ms":421,"total_runtime_ms":2362,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0295,"confidence_interval":0.9401,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6894,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2127,"p95_latency_ms":2599,"p99_latency_ms":2788,"time_to_first_token_ms":440,"total_runtime_ms":2363,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0269,"confidence_interval":0.9432,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6199,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2050,"p95_latency_ms":2506,"p99_latency_ms":2688,"time_to_first_token_ms":435,"total_runtime_ms":2278,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0117,"confidence_interval":0.9471,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0718,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2090,"p95_latency_ms":2554,"p99_latency_ms":2740,"time_to_first_token_ms":436,"total_runtime_ms":2322,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0262,"confidence_interval":0.9491,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.44,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2126,"p95_latency_ms":2599,"p99_latency_ms":2788,"time_to_first_token_ms":431,"total_runtime_ms":2363,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.021,"confidence_interval":0.9353,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1328,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2044,"p95_latency_ms":2498,"p99_latency_ms":2679,"time_to_first_token_ms":423,"total_runtime_ms":2271,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0198,"confidence_interval":0.935,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5316,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2090,"p95_latency_ms":2554,"p99_latency_ms":2740,"time_to_first_token_ms":431,"total_runtime_ms":2322,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0297,"confidence_interval":0.9338,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7995,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1095,"p95_latency_ms":1338,"p99_latency_ms":1435,"time_to_first_token_ms":390,"total_runtime_ms":1216,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0248,"confidence_interval":0.9454,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6806,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1091,"p95_latency_ms":1334,"p99_latency_ms":1431,"time_to_first_token_ms":384,"total_runtime_ms":1212,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0171,"confidence_interval":0.9394,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7354,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1073,"p95_latency_ms":1312,"p99_latency_ms":1407,"time_to_first_token_ms":373,"total_runtime_ms":1193,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0255,"confidence_interval":0.9496,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4951,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1090,"p95_latency_ms":1332,"p99_latency_ms":1429,"time_to_first_token_ms":384,"total_runtime_ms":1211,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0187,"confidence_interval":0.9493,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0326,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1040,"p95_latency_ms":1271,"p99_latency_ms":1363,"time_to_first_token_ms":386,"total_runtime_ms":1155,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0148,"confidence_interval":0.9452,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7541,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1116,"p95_latency_ms":1364,"p99_latency_ms":1463,"time_to_first_token_ms":382,"total_runtime_ms":1240,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0119,"confidence_interval":0.9387,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0731,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1087,"p95_latency_ms":1329,"p99_latency_ms":1426,"time_to_first_token_ms":382,"total_runtime_ms":1208,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0102,"confidence_interval":0.9485,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2507,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1059,"p95_latency_ms":1294,"p99_latency_ms":1388,"time_to_first_token_ms":383,"total_runtime_ms":1177,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0169,"confidence_interval":0.9376,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4389,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1018,"p95_latency_ms":1245,"p99_latency_ms":1335,"time_to_first_token_ms":392,"total_runtime_ms":1132,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0207,"confidence_interval":0.9381,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5262,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:7b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1065,"p95_latency_ms":1302,"p99_latency_ms":1397,"time_to_first_token_ms":370,"total_runtime_ms":1184,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0255,"confidence_interval":0.9387,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2292,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3092,"p95_latency_ms":3779,"p99_latency_ms":4054,"time_to_first_token_ms":575,"total_runtime_ms":3436,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0177,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8014,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3077,"p95_latency_ms":3761,"p99_latency_ms":4034,"time_to_first_token_ms":599,"total_runtime_ms":3419,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0221,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3467,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3063,"p95_latency_ms":3743,"p99_latency_ms":4016,"time_to_first_token_ms":600,"total_runtime_ms":3403,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0226,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6312,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3015,"p95_latency_ms":3685,"p99_latency_ms":3953,"time_to_first_token_ms":593,"total_runtime_ms":3350,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.018,"confidence_interval":0.9349,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7306,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3031,"p95_latency_ms":3704,"p99_latency_ms":3974,"time_to_first_token_ms":582,"total_runtime_ms":3368,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0178,"confidence_interval":0.9391,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6771,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3021,"p95_latency_ms":3692,"p99_latency_ms":3960,"time_to_first_token_ms":603,"total_runtime_ms":3356,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0241,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8966,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3056,"p95_latency_ms":3736,"p99_latency_ms":4007,"time_to_first_token_ms":580,"total_runtime_ms":3396,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0113,"confidence_interval":0.9337,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.754,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3002,"p95_latency_ms":3669,"p99_latency_ms":3936,"time_to_first_token_ms":588,"total_runtime_ms":3335,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0269,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1268,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3051,"p95_latency_ms":3729,"p99_latency_ms":4000,"time_to_first_token_ms":589,"total_runtime_ms":3390,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0187,"confidence_interval":0.9494,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2404,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3014,"p95_latency_ms":3684,"p99_latency_ms":3952,"time_to_first_token_ms":587,"total_runtime_ms":3349,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0168,"confidence_interval":0.9312,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6847,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1521,"p95_latency_ms":1859,"p99_latency_ms":1995,"time_to_first_token_ms":546,"total_runtime_ms":1690,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0212,"confidence_interval":0.9324,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1988,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1554,"p95_latency_ms":1899,"p99_latency_ms":2037,"time_to_first_token_ms":540,"total_runtime_ms":1726,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.949,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3602,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1568,"p95_latency_ms":1916,"p99_latency_ms":2056,"time_to_first_token_ms":562,"total_runtime_ms":1742,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0273,"confidence_interval":0.9347,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0683,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1572,"p95_latency_ms":1922,"p99_latency_ms":2061,"time_to_first_token_ms":541,"total_runtime_ms":1747,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0299,"confidence_interval":0.9489,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7512,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1539,"p95_latency_ms":1881,"p99_latency_ms":2018,"time_to_first_token_ms":561,"total_runtime_ms":1710,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0155,"confidence_interval":0.9377,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3882,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1512,"p95_latency_ms":1848,"p99_latency_ms":1982,"time_to_first_token_ms":531,"total_runtime_ms":1680,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6321,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1600,"p95_latency_ms":1955,"p99_latency_ms":2097,"time_to_first_token_ms":553,"total_runtime_ms":1777,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0185,"confidence_interval":0.9339,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7487,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1506,"p95_latency_ms":1841,"p99_latency_ms":1975,"time_to_first_token_ms":546,"total_runtime_ms":1674,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0137,"confidence_interval":0.943,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8841,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1516,"p95_latency_ms":1853,"p99_latency_ms":1988,"time_to_first_token_ms":535,"total_runtime_ms":1684,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0182,"confidence_interval":0.9304,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5011,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|low|qwen2:14b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1543,"p95_latency_ms":1886,"p99_latency_ms":2023,"time_to_first_token_ms":536,"total_runtime_ms":1714,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0293,"confidence_interval":0.9407,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9573,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":638,"p95_latency_ms":780,"p99_latency_ms":837,"time_to_first_token_ms":186,"total_runtime_ms":709,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3063,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":613,"p95_latency_ms":749,"p99_latency_ms":803,"time_to_first_token_ms":159,"total_runtime_ms":681,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0123,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.394,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":632,"p95_latency_ms":773,"p99_latency_ms":829,"time_to_first_token_ms":189,"total_runtime_ms":703,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0232,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9562,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":607,"p95_latency_ms":742,"p99_latency_ms":796,"time_to_first_token_ms":184,"total_runtime_ms":674,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0112,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2245,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":665,"p95_latency_ms":813,"p99_latency_ms":872,"time_to_first_token_ms":159,"total_runtime_ms":739,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0211,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6001,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":186,"total_runtime_ms":664,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0249,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5336,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":626,"p95_latency_ms":765,"p99_latency_ms":821,"time_to_first_token_ms":165,"total_runtime_ms":696,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0133,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.55,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":609,"p95_latency_ms":744,"p99_latency_ms":798,"time_to_first_token_ms":178,"total_runtime_ms":677,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0228,"confidence_interval":0.9397,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6653,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":168,"total_runtime_ms":665,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0246,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5225,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":743,"time_to_first_token_ms":158,"total_runtime_ms":630,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0142,"confidence_interval":0.9357,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9585,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":361,"p95_latency_ms":441,"p99_latency_ms":473,"time_to_first_token_ms":113,"total_runtime_ms":401,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0193,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8403,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":314,"p95_latency_ms":384,"p99_latency_ms":412,"time_to_first_token_ms":134,"total_runtime_ms":349,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0103,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9422,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":438,"p99_latency_ms":469,"time_to_first_token_ms":134,"total_runtime_ms":398,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0154,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0132,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":357,"p95_latency_ms":436,"p99_latency_ms":468,"time_to_first_token_ms":124,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.025,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2323,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":346,"p95_latency_ms":423,"p99_latency_ms":454,"time_to_first_token_ms":116,"total_runtime_ms":385,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0143,"confidence_interval":0.9395,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5278,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":286,"p95_latency_ms":349,"p99_latency_ms":375,"time_to_first_token_ms":139,"total_runtime_ms":318,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0109,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3461,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":506,"time_to_first_token_ms":134,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0209,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4055,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":305,"p95_latency_ms":373,"p99_latency_ms":400,"time_to_first_token_ms":122,"total_runtime_ms":339,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0258,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2307,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":376,"p99_latency_ms":404,"time_to_first_token_ms":140,"total_runtime_ms":342,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0128,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5853,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":360,"p95_latency_ms":440,"p99_latency_ms":472,"time_to_first_token_ms":139,"total_runtime_ms":400,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0137,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5828,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":844,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":200,"total_runtime_ms":938,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0213,"confidence_interval":0.9487,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.142,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":856,"p95_latency_ms":1047,"p99_latency_ms":1123,"time_to_first_token_ms":224,"total_runtime_ms":951,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0101,"confidence_interval":0.9352,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3567,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":761,"p95_latency_ms":930,"p99_latency_ms":997,"time_to_first_token_ms":203,"total_runtime_ms":845,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0288,"confidence_interval":0.934,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6482,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":777,"p95_latency_ms":950,"p99_latency_ms":1019,"time_to_first_token_ms":208,"total_runtime_ms":863,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0103,"confidence_interval":0.9431,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9784,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":829,"p95_latency_ms":1013,"p99_latency_ms":1086,"time_to_first_token_ms":197,"total_runtime_ms":921,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0293,"confidence_interval":0.9372,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0338,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":852,"p95_latency_ms":1041,"p99_latency_ms":1117,"time_to_first_token_ms":206,"total_runtime_ms":946,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0204,"confidence_interval":0.9402,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1774,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":784,"p95_latency_ms":958,"p99_latency_ms":1027,"time_to_first_token_ms":221,"total_runtime_ms":871,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.022,"confidence_interval":0.9337,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5274,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":851,"p95_latency_ms":1040,"p99_latency_ms":1116,"time_to_first_token_ms":213,"total_runtime_ms":945,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0256,"confidence_interval":0.9394,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8237,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":796,"p95_latency_ms":973,"p99_latency_ms":1043,"time_to_first_token_ms":218,"total_runtime_ms":884,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0105,"confidence_interval":0.9458,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3788,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":763,"p95_latency_ms":932,"p99_latency_ms":1000,"time_to_first_token_ms":223,"total_runtime_ms":847,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0192,"confidence_interval":0.9475,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9656,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":412,"p95_latency_ms":503,"p99_latency_ms":540,"time_to_first_token_ms":149,"total_runtime_ms":458,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0128,"confidence_interval":0.9381,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0615,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":478,"p95_latency_ms":585,"p99_latency_ms":627,"time_to_first_token_ms":157,"total_runtime_ms":531,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0119,"confidence_interval":0.9409,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5006,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":426,"p95_latency_ms":521,"p99_latency_ms":559,"time_to_first_token_ms":170,"total_runtime_ms":474,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0146,"confidence_interval":0.9348,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1736,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":161,"total_runtime_ms":504,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0133,"confidence_interval":0.9322,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.349,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":457,"p95_latency_ms":559,"p99_latency_ms":599,"time_to_first_token_ms":165,"total_runtime_ms":508,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0268,"confidence_interval":0.9445,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3102,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":396,"p95_latency_ms":484,"p99_latency_ms":519,"time_to_first_token_ms":165,"total_runtime_ms":440,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0148,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9856,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":422,"p95_latency_ms":516,"p99_latency_ms":554,"time_to_first_token_ms":147,"total_runtime_ms":469,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0255,"confidence_interval":0.9319,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3038,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":442,"p95_latency_ms":540,"p99_latency_ms":580,"time_to_first_token_ms":169,"total_runtime_ms":491,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0273,"confidence_interval":0.9343,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8917,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":412,"p95_latency_ms":504,"p99_latency_ms":541,"time_to_first_token_ms":146,"total_runtime_ms":458,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0109,"confidence_interval":0.9419,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5091,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":403,"p95_latency_ms":493,"p99_latency_ms":529,"time_to_first_token_ms":162,"total_runtime_ms":448,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0197,"confidence_interval":0.9387,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7873,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1567,"p95_latency_ms":1915,"p99_latency_ms":2055,"time_to_first_token_ms":325,"total_runtime_ms":1741,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0206,"confidence_interval":0.937,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1935,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1558,"p95_latency_ms":1904,"p99_latency_ms":2043,"time_to_first_token_ms":323,"total_runtime_ms":1731,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.03,"confidence_interval":0.9467,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8197,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1622,"p95_latency_ms":1983,"p99_latency_ms":2127,"time_to_first_token_ms":337,"total_runtime_ms":1803,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0157,"confidence_interval":0.9345,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3331,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1588,"p95_latency_ms":1941,"p99_latency_ms":2083,"time_to_first_token_ms":334,"total_runtime_ms":1765,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0273,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6396,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1558,"p95_latency_ms":1905,"p99_latency_ms":2043,"time_to_first_token_ms":332,"total_runtime_ms":1731,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0237,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6044,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1604,"p95_latency_ms":1960,"p99_latency_ms":2102,"time_to_first_token_ms":349,"total_runtime_ms":1782,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0146,"confidence_interval":0.9309,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0002,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1595,"p95_latency_ms":1949,"p99_latency_ms":2091,"time_to_first_token_ms":344,"total_runtime_ms":1772,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0104,"confidence_interval":0.9433,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9609,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1575,"p95_latency_ms":1926,"p99_latency_ms":2066,"time_to_first_token_ms":323,"total_runtime_ms":1750,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0162,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9323,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1535,"p95_latency_ms":1876,"p99_latency_ms":2013,"time_to_first_token_ms":333,"total_runtime_ms":1706,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0159,"confidence_interval":0.9472,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4638,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1589,"p95_latency_ms":1942,"p99_latency_ms":2083,"time_to_first_token_ms":320,"total_runtime_ms":1765,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0191,"confidence_interval":0.9379,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8375,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":775,"p95_latency_ms":947,"p99_latency_ms":1016,"time_to_first_token_ms":299,"total_runtime_ms":861,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0107,"confidence_interval":0.934,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7313,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":784,"p95_latency_ms":958,"p99_latency_ms":1028,"time_to_first_token_ms":294,"total_runtime_ms":871,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0292,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6747,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1052,"time_to_first_token_ms":296,"total_runtime_ms":892,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0126,"confidence_interval":0.9373,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9342,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1010,"time_to_first_token_ms":287,"total_runtime_ms":856,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0119,"confidence_interval":0.9493,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4797,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":816,"p95_latency_ms":997,"p99_latency_ms":1069,"time_to_first_token_ms":300,"total_runtime_ms":906,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0111,"confidence_interval":0.9301,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7659,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":779,"p95_latency_ms":953,"p99_latency_ms":1022,"time_to_first_token_ms":284,"total_runtime_ms":866,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0107,"confidence_interval":0.9454,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0601,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":825,"p95_latency_ms":1008,"p99_latency_ms":1082,"time_to_first_token_ms":280,"total_runtime_ms":917,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0142,"confidence_interval":0.9475,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3432,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":293,"total_runtime_ms":895,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0149,"confidence_interval":0.942,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0363,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":839,"p95_latency_ms":1026,"p99_latency_ms":1100,"time_to_first_token_ms":282,"total_runtime_ms":933,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0107,"confidence_interval":0.9331,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2066,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":789,"p95_latency_ms":965,"p99_latency_ms":1035,"time_to_first_token_ms":276,"total_runtime_ms":877,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0168,"confidence_interval":0.937,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0204,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2256,"p95_latency_ms":2758,"p99_latency_ms":2958,"time_to_first_token_ms":463,"total_runtime_ms":2507,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0266,"confidence_interval":0.9341,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8376,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2315,"p95_latency_ms":2830,"p99_latency_ms":3035,"time_to_first_token_ms":460,"total_runtime_ms":2572,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0184,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1284,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2256,"p95_latency_ms":2758,"p99_latency_ms":2958,"time_to_first_token_ms":457,"total_runtime_ms":2507,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0231,"confidence_interval":0.9361,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3261,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2301,"p95_latency_ms":2813,"p99_latency_ms":3017,"time_to_first_token_ms":469,"total_runtime_ms":2557,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0111,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1646,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2334,"p95_latency_ms":2853,"p99_latency_ms":3060,"time_to_first_token_ms":463,"total_runtime_ms":2593,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0175,"confidence_interval":0.945,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5329,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2301,"p95_latency_ms":2812,"p99_latency_ms":3017,"time_to_first_token_ms":463,"total_runtime_ms":2557,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0269,"confidence_interval":0.9327,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1015,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2300,"p95_latency_ms":2811,"p99_latency_ms":3015,"time_to_first_token_ms":452,"total_runtime_ms":2555,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0148,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0115,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2281,"p95_latency_ms":2788,"p99_latency_ms":2991,"time_to_first_token_ms":468,"total_runtime_ms":2535,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0111,"confidence_interval":0.9318,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9888,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2344,"p95_latency_ms":2865,"p99_latency_ms":3074,"time_to_first_token_ms":471,"total_runtime_ms":2605,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0286,"confidence_interval":0.9396,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2541,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2349,"p95_latency_ms":2871,"p99_latency_ms":3080,"time_to_first_token_ms":472,"total_runtime_ms":2610,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0233,"confidence_interval":0.9306,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2708,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1178,"p95_latency_ms":1440,"p99_latency_ms":1545,"time_to_first_token_ms":420,"total_runtime_ms":1309,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.019,"confidence_interval":0.9339,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1821,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1230,"p95_latency_ms":1503,"p99_latency_ms":1612,"time_to_first_token_ms":409,"total_runtime_ms":1366,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0269,"confidence_interval":0.9483,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.923,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1204,"p95_latency_ms":1472,"p99_latency_ms":1579,"time_to_first_token_ms":407,"total_runtime_ms":1338,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0112,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.287,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1147,"p95_latency_ms":1401,"p99_latency_ms":1503,"time_to_first_token_ms":410,"total_runtime_ms":1274,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0161,"confidence_interval":0.9445,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6232,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1136,"p95_latency_ms":1389,"p99_latency_ms":1489,"time_to_first_token_ms":427,"total_runtime_ms":1262,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0264,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8954,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1180,"p95_latency_ms":1442,"p99_latency_ms":1547,"time_to_first_token_ms":423,"total_runtime_ms":1311,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0263,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2087,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1176,"p95_latency_ms":1437,"p99_latency_ms":1542,"time_to_first_token_ms":434,"total_runtime_ms":1307,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0166,"confidence_interval":0.939,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1238,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1162,"p95_latency_ms":1421,"p99_latency_ms":1524,"time_to_first_token_ms":413,"total_runtime_ms":1292,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0198,"confidence_interval":0.934,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3336,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1155,"p95_latency_ms":1411,"p99_latency_ms":1514,"time_to_first_token_ms":415,"total_runtime_ms":1283,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0157,"confidence_interval":0.9332,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1143,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1147,"p95_latency_ms":1401,"p99_latency_ms":1503,"time_to_first_token_ms":425,"total_runtime_ms":1274,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.026,"confidence_interval":0.9368,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5915,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":147,"total_runtime_ms":505,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9348,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7851,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":431,"p95_latency_ms":527,"p99_latency_ms":565,"time_to_first_token_ms":149,"total_runtime_ms":479,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0125,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4155,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":513,"p95_latency_ms":627,"p99_latency_ms":673,"time_to_first_token_ms":153,"total_runtime_ms":570,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0236,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0284,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":480,"p95_latency_ms":586,"p99_latency_ms":629,"time_to_first_token_ms":156,"total_runtime_ms":533,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0248,"confidence_interval":0.9361,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0901,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":494,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":165,"total_runtime_ms":549,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0239,"confidence_interval":0.9468,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0403,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":434,"p95_latency_ms":530,"p99_latency_ms":569,"time_to_first_token_ms":157,"total_runtime_ms":482,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0116,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.63,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":452,"p95_latency_ms":552,"p99_latency_ms":593,"time_to_first_token_ms":139,"total_runtime_ms":502,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0158,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1793,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":509,"p95_latency_ms":622,"p99_latency_ms":667,"time_to_first_token_ms":163,"total_runtime_ms":565,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0229,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7354,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":495,"p95_latency_ms":605,"p99_latency_ms":649,"time_to_first_token_ms":147,"total_runtime_ms":550,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0284,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0197,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":493,"p95_latency_ms":602,"p99_latency_ms":646,"time_to_first_token_ms":159,"total_runtime_ms":548,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.013,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9471,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":252,"p95_latency_ms":307,"p99_latency_ms":330,"time_to_first_token_ms":91,"total_runtime_ms":280,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5007,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":244,"p95_latency_ms":298,"p99_latency_ms":320,"time_to_first_token_ms":91,"total_runtime_ms":271,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0262,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8007,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":289,"p95_latency_ms":354,"p99_latency_ms":379,"time_to_first_token_ms":101,"total_runtime_ms":321,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0165,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0272,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":263,"p95_latency_ms":321,"p99_latency_ms":344,"time_to_first_token_ms":102,"total_runtime_ms":292,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0231,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9866,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":290,"p99_latency_ms":312,"time_to_first_token_ms":89,"total_runtime_ms":264,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0282,"confidence_interval":0.9317,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1247,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":317,"p95_latency_ms":387,"p99_latency_ms":415,"time_to_first_token_ms":98,"total_runtime_ms":352,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0181,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.729,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":215,"p95_latency_ms":262,"p99_latency_ms":281,"time_to_first_token_ms":107,"total_runtime_ms":239,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0104,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2451,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":225,"p95_latency_ms":275,"p99_latency_ms":295,"time_to_first_token_ms":113,"total_runtime_ms":250,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0282,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9788,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":248,"p95_latency_ms":304,"p99_latency_ms":326,"time_to_first_token_ms":122,"total_runtime_ms":276,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0125,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.328,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":261,"p95_latency_ms":319,"p99_latency_ms":342,"time_to_first_token_ms":115,"total_runtime_ms":290,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0223,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.034,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":635,"p95_latency_ms":776,"p99_latency_ms":832,"time_to_first_token_ms":157,"total_runtime_ms":705,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.027,"confidence_interval":0.9351,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7934,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":613,"p95_latency_ms":750,"p99_latency_ms":804,"time_to_first_token_ms":177,"total_runtime_ms":682,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.9424,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6873,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":653,"p95_latency_ms":798,"p99_latency_ms":856,"time_to_first_token_ms":165,"total_runtime_ms":725,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.018,"confidence_interval":0.9352,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4034,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":661,"p95_latency_ms":807,"p99_latency_ms":866,"time_to_first_token_ms":182,"total_runtime_ms":734,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0226,"confidence_interval":0.9463,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0536,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":616,"p95_latency_ms":752,"p99_latency_ms":807,"time_to_first_token_ms":186,"total_runtime_ms":684,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.028,"confidence_interval":0.9345,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.908,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":598,"p95_latency_ms":730,"p99_latency_ms":784,"time_to_first_token_ms":188,"total_runtime_ms":664,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0239,"confidence_interval":0.948,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9057,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":660,"p95_latency_ms":806,"p99_latency_ms":865,"time_to_first_token_ms":180,"total_runtime_ms":733,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.011,"confidence_interval":0.9482,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0958,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":591,"p95_latency_ms":723,"p99_latency_ms":775,"time_to_first_token_ms":178,"total_runtime_ms":657,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0229,"confidence_interval":0.9472,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5001,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":616,"p95_latency_ms":753,"p99_latency_ms":808,"time_to_first_token_ms":189,"total_runtime_ms":685,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.9365,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8025,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":803,"time_to_first_token_ms":171,"total_runtime_ms":680,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0286,"confidence_interval":0.9429,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6585,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":408,"time_to_first_token_ms":139,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2163,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":376,"p99_latency_ms":403,"time_to_first_token_ms":147,"total_runtime_ms":342,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0247,"confidence_interval":0.9397,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.039,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":335,"p95_latency_ms":410,"p99_latency_ms":440,"time_to_first_token_ms":133,"total_runtime_ms":373,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0205,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4025,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":323,"p95_latency_ms":395,"p99_latency_ms":424,"time_to_first_token_ms":114,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0186,"confidence_interval":0.9452,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3253,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":304,"p95_latency_ms":372,"p99_latency_ms":399,"time_to_first_token_ms":133,"total_runtime_ms":338,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0121,"confidence_interval":0.9467,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.705,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":113,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0136,"confidence_interval":0.932,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8069,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":306,"p95_latency_ms":374,"p99_latency_ms":401,"time_to_first_token_ms":118,"total_runtime_ms":340,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0284,"confidence_interval":0.9406,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2511,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":367,"p95_latency_ms":448,"p99_latency_ms":481,"time_to_first_token_ms":120,"total_runtime_ms":408,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0124,"confidence_interval":0.9457,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.69,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":327,"p95_latency_ms":400,"p99_latency_ms":429,"time_to_first_token_ms":114,"total_runtime_ms":363,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0284,"confidence_interval":0.9493,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8365,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":362,"p95_latency_ms":442,"p99_latency_ms":474,"time_to_first_token_ms":144,"total_runtime_ms":402,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0246,"confidence_interval":0.9462,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2245,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1140,"p95_latency_ms":1393,"p99_latency_ms":1495,"time_to_first_token_ms":278,"total_runtime_ms":1267,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0188,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4169,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1161,"p95_latency_ms":1419,"p99_latency_ms":1523,"time_to_first_token_ms":276,"total_runtime_ms":1290,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0122,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2163,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1222,"p95_latency_ms":1493,"p99_latency_ms":1602,"time_to_first_token_ms":285,"total_runtime_ms":1357,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0205,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4292,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1172,"p95_latency_ms":1433,"p99_latency_ms":1537,"time_to_first_token_ms":273,"total_runtime_ms":1302,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0211,"confidence_interval":0.9492,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5243,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1205,"p95_latency_ms":1473,"p99_latency_ms":1580,"time_to_first_token_ms":258,"total_runtime_ms":1339,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0111,"confidence_interval":0.944,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2447,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1172,"p95_latency_ms":1432,"p99_latency_ms":1536,"time_to_first_token_ms":270,"total_runtime_ms":1302,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0152,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1549,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1187,"p95_latency_ms":1451,"p99_latency_ms":1556,"time_to_first_token_ms":280,"total_runtime_ms":1319,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0143,"confidence_interval":0.933,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8828,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1223,"p95_latency_ms":1495,"p99_latency_ms":1603,"time_to_first_token_ms":273,"total_runtime_ms":1359,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0161,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1795,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1231,"p95_latency_ms":1504,"p99_latency_ms":1614,"time_to_first_token_ms":285,"total_runtime_ms":1367,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0253,"confidence_interval":0.9329,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6625,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1184,"p95_latency_ms":1448,"p99_latency_ms":1553,"time_to_first_token_ms":275,"total_runtime_ms":1316,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0122,"confidence_interval":0.9416,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8692,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":669,"p95_latency_ms":818,"p99_latency_ms":877,"time_to_first_token_ms":226,"total_runtime_ms":743,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0232,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1692,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":612,"p95_latency_ms":747,"p99_latency_ms":802,"time_to_first_token_ms":242,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0175,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8368,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":646,"p95_latency_ms":789,"p99_latency_ms":847,"time_to_first_token_ms":214,"total_runtime_ms":718,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0156,"confidence_interval":0.9306,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1916,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":665,"p95_latency_ms":813,"p99_latency_ms":872,"time_to_first_token_ms":216,"total_runtime_ms":739,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7868,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":636,"p95_latency_ms":777,"p99_latency_ms":834,"time_to_first_token_ms":221,"total_runtime_ms":706,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0134,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6033,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":609,"p95_latency_ms":744,"p99_latency_ms":798,"time_to_first_token_ms":215,"total_runtime_ms":676,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0283,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3488,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":749,"time_to_first_token_ms":233,"total_runtime_ms":635,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0195,"confidence_interval":0.9323,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8943,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":570,"p95_latency_ms":697,"p99_latency_ms":748,"time_to_first_token_ms":232,"total_runtime_ms":634,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.022,"confidence_interval":0.941,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7224,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":570,"p95_latency_ms":697,"p99_latency_ms":747,"time_to_first_token_ms":233,"total_runtime_ms":633,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0251,"confidence_interval":0.9458,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1612,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":843,"time_to_first_token_ms":212,"total_runtime_ms":715,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0232,"confidence_interval":0.9342,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1157,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1727,"p95_latency_ms":2111,"p99_latency_ms":2265,"time_to_first_token_ms":374,"total_runtime_ms":1919,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.946,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9086,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1780,"p95_latency_ms":2176,"p99_latency_ms":2334,"time_to_first_token_ms":355,"total_runtime_ms":1978,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8376,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1786,"p95_latency_ms":2182,"p99_latency_ms":2341,"time_to_first_token_ms":371,"total_runtime_ms":1984,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0216,"confidence_interval":0.9355,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4277,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1689,"p95_latency_ms":2064,"p99_latency_ms":2214,"time_to_first_token_ms":372,"total_runtime_ms":1876,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0124,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2892,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1771,"p95_latency_ms":2165,"p99_latency_ms":2322,"time_to_first_token_ms":378,"total_runtime_ms":1968,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0143,"confidence_interval":0.9456,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7705,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1758,"p95_latency_ms":2148,"p99_latency_ms":2305,"time_to_first_token_ms":373,"total_runtime_ms":1953,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.016,"confidence_interval":0.9439,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9541,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1715,"p95_latency_ms":2096,"p99_latency_ms":2248,"time_to_first_token_ms":350,"total_runtime_ms":1905,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0155,"confidence_interval":0.9472,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8508,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1773,"p95_latency_ms":2167,"p99_latency_ms":2324,"time_to_first_token_ms":349,"total_runtime_ms":1970,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.026,"confidence_interval":0.9398,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9432,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1689,"p95_latency_ms":2064,"p99_latency_ms":2214,"time_to_first_token_ms":361,"total_runtime_ms":1877,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0173,"confidence_interval":0.936,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.31,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1687,"p95_latency_ms":2062,"p99_latency_ms":2212,"time_to_first_token_ms":353,"total_runtime_ms":1874,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0161,"confidence_interval":0.9422,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8498,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":900,"p95_latency_ms":1100,"p99_latency_ms":1180,"time_to_first_token_ms":307,"total_runtime_ms":1000,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0237,"confidence_interval":0.9492,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7687,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":913,"p95_latency_ms":1116,"p99_latency_ms":1197,"time_to_first_token_ms":325,"total_runtime_ms":1014,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0203,"confidence_interval":0.9313,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6678,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":879,"p95_latency_ms":1074,"p99_latency_ms":1152,"time_to_first_token_ms":311,"total_runtime_ms":977,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0102,"confidence_interval":0.9324,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1057,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":888,"p95_latency_ms":1086,"p99_latency_ms":1165,"time_to_first_token_ms":335,"total_runtime_ms":987,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0267,"confidence_interval":0.9368,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1441,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":927,"p95_latency_ms":1133,"p99_latency_ms":1215,"time_to_first_token_ms":328,"total_runtime_ms":1030,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0299,"confidence_interval":0.9449,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9101,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":940,"p95_latency_ms":1149,"p99_latency_ms":1232,"time_to_first_token_ms":337,"total_runtime_ms":1044,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0158,"confidence_interval":0.9354,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.294,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":934,"p95_latency_ms":1141,"p99_latency_ms":1224,"time_to_first_token_ms":312,"total_runtime_ms":1037,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0115,"confidence_interval":0.9349,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4301,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":918,"p95_latency_ms":1122,"p99_latency_ms":1204,"time_to_first_token_ms":318,"total_runtime_ms":1020,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0292,"confidence_interval":0.9375,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6062,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":895,"p95_latency_ms":1094,"p99_latency_ms":1174,"time_to_first_token_ms":308,"total_runtime_ms":995,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0182,"confidence_interval":0.9365,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5278,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":911,"p95_latency_ms":1113,"p99_latency_ms":1194,"time_to_first_token_ms":311,"total_runtime_ms":1012,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0136,"confidence_interval":0.9418,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.903,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":432,"p95_latency_ms":528,"p99_latency_ms":567,"time_to_first_token_ms":143,"total_runtime_ms":480,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0117,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6319,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":596,"time_to_first_token_ms":152,"total_runtime_ms":505,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0244,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0963,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":436,"p95_latency_ms":533,"p99_latency_ms":571,"time_to_first_token_ms":152,"total_runtime_ms":484,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0249,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4178,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":357,"p95_latency_ms":436,"p99_latency_ms":467,"time_to_first_token_ms":148,"total_runtime_ms":396,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0142,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7732,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":367,"p95_latency_ms":449,"p99_latency_ms":482,"time_to_first_token_ms":131,"total_runtime_ms":408,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0163,"confidence_interval":0.9402,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1357,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":450,"p95_latency_ms":550,"p99_latency_ms":590,"time_to_first_token_ms":150,"total_runtime_ms":500,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0266,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7193,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":371,"p95_latency_ms":453,"p99_latency_ms":486,"time_to_first_token_ms":152,"total_runtime_ms":412,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0204,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6029,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":385,"p95_latency_ms":470,"p99_latency_ms":505,"time_to_first_token_ms":148,"total_runtime_ms":428,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0204,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6121,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":367,"p95_latency_ms":448,"p99_latency_ms":481,"time_to_first_token_ms":153,"total_runtime_ms":407,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0286,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8441,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":355,"p95_latency_ms":434,"p99_latency_ms":465,"time_to_first_token_ms":151,"total_runtime_ms":394,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0284,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6439,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":195,"p95_latency_ms":238,"p99_latency_ms":255,"time_to_first_token_ms":80,"total_runtime_ms":216,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0153,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5439,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":184,"p95_latency_ms":225,"p99_latency_ms":241,"time_to_first_token_ms":78,"total_runtime_ms":204,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1749,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":263,"p95_latency_ms":322,"p99_latency_ms":345,"time_to_first_token_ms":77,"total_runtime_ms":293,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0151,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8028,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":222,"p95_latency_ms":271,"p99_latency_ms":290,"time_to_first_token_ms":105,"total_runtime_ms":246,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0185,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3939,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":201,"p95_latency_ms":245,"p99_latency_ms":263,"time_to_first_token_ms":99,"total_runtime_ms":223,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0285,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3825,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":259,"p95_latency_ms":317,"p99_latency_ms":340,"time_to_first_token_ms":108,"total_runtime_ms":288,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0105,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.415,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":260,"p95_latency_ms":318,"p99_latency_ms":341,"time_to_first_token_ms":107,"total_runtime_ms":289,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0299,"confidence_interval":0.9451,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5672,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":219,"p95_latency_ms":268,"p99_latency_ms":288,"time_to_first_token_ms":85,"total_runtime_ms":244,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0284,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3524,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":192,"p95_latency_ms":235,"p99_latency_ms":252,"time_to_first_token_ms":96,"total_runtime_ms":214,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0278,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5982,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":266,"p95_latency_ms":325,"p99_latency_ms":348,"time_to_first_token_ms":102,"total_runtime_ms":295,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0227,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2816,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":744,"time_to_first_token_ms":165,"total_runtime_ms":630,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0203,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.474,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":546,"p95_latency_ms":667,"p99_latency_ms":715,"time_to_first_token_ms":143,"total_runtime_ms":606,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.027,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7458,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":690,"time_to_first_token_ms":159,"total_runtime_ms":585,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0163,"confidence_interval":0.9338,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5508,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":501,"p95_latency_ms":613,"p99_latency_ms":657,"time_to_first_token_ms":152,"total_runtime_ms":557,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0116,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.094,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":148,"total_runtime_ms":611,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0222,"confidence_interval":0.9332,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4118,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":530,"p95_latency_ms":648,"p99_latency_ms":695,"time_to_first_token_ms":145,"total_runtime_ms":589,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0161,"confidence_interval":0.9475,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9241,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":152,"total_runtime_ms":636,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.01,"confidence_interval":0.9307,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8501,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":565,"p95_latency_ms":691,"p99_latency_ms":741,"time_to_first_token_ms":154,"total_runtime_ms":628,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0187,"confidence_interval":0.9466,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.925,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":158,"total_runtime_ms":636,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0288,"confidence_interval":0.95,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5017,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":698,"time_to_first_token_ms":154,"total_runtime_ms":591,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0227,"confidence_interval":0.9361,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.852,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":241,"p95_latency_ms":295,"p99_latency_ms":317,"time_to_first_token_ms":102,"total_runtime_ms":268,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0162,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1186,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":239,"p95_latency_ms":293,"p99_latency_ms":314,"time_to_first_token_ms":107,"total_runtime_ms":266,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.9306,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3309,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":271,"p95_latency_ms":331,"p99_latency_ms":355,"time_to_first_token_ms":117,"total_runtime_ms":301,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0282,"confidence_interval":0.9499,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0874,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":286,"p95_latency_ms":350,"p99_latency_ms":375,"time_to_first_token_ms":128,"total_runtime_ms":318,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.9423,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.359,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":341,"p95_latency_ms":417,"p99_latency_ms":447,"time_to_first_token_ms":124,"total_runtime_ms":379,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0186,"confidence_interval":0.9464,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6393,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":298,"p95_latency_ms":365,"p99_latency_ms":391,"time_to_first_token_ms":117,"total_runtime_ms":332,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0173,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5702,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":285,"p95_latency_ms":349,"p99_latency_ms":374,"time_to_first_token_ms":105,"total_runtime_ms":317,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0176,"confidence_interval":0.9382,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4554,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":327,"p95_latency_ms":399,"p99_latency_ms":428,"time_to_first_token_ms":99,"total_runtime_ms":363,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0119,"confidence_interval":0.9314,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4024,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":300,"p95_latency_ms":367,"p99_latency_ms":393,"time_to_first_token_ms":107,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0289,"confidence_interval":0.9439,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9638,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":425,"time_to_first_token_ms":130,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0239,"confidence_interval":0.9361,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3279,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1035,"p95_latency_ms":1265,"p99_latency_ms":1357,"time_to_first_token_ms":231,"total_runtime_ms":1150,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0217,"confidence_interval":0.9301,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4347,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":980,"p95_latency_ms":1197,"p99_latency_ms":1285,"time_to_first_token_ms":241,"total_runtime_ms":1089,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0277,"confidence_interval":0.9449,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6516,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1051,"p95_latency_ms":1284,"p99_latency_ms":1377,"time_to_first_token_ms":236,"total_runtime_ms":1167,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0181,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4065,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1032,"p95_latency_ms":1262,"p99_latency_ms":1353,"time_to_first_token_ms":225,"total_runtime_ms":1147,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.9479,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2191,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":989,"p95_latency_ms":1209,"p99_latency_ms":1296,"time_to_first_token_ms":225,"total_runtime_ms":1099,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0271,"confidence_interval":0.9374,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2305,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1042,"p95_latency_ms":1273,"p99_latency_ms":1366,"time_to_first_token_ms":222,"total_runtime_ms":1158,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0111,"confidence_interval":0.938,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9234,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":948,"p95_latency_ms":1158,"p99_latency_ms":1243,"time_to_first_token_ms":241,"total_runtime_ms":1053,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0122,"confidence_interval":0.9395,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.536,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":963,"p95_latency_ms":1177,"p99_latency_ms":1262,"time_to_first_token_ms":235,"total_runtime_ms":1070,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0249,"confidence_interval":0.9431,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6672,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1048,"p95_latency_ms":1280,"p99_latency_ms":1374,"time_to_first_token_ms":245,"total_runtime_ms":1164,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0247,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0781,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":977,"p95_latency_ms":1194,"p99_latency_ms":1281,"time_to_first_token_ms":257,"total_runtime_ms":1085,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0103,"confidence_interval":0.9415,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5773,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":486,"p95_latency_ms":595,"p99_latency_ms":638,"time_to_first_token_ms":182,"total_runtime_ms":540,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0108,"confidence_interval":0.938,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2011,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":756,"time_to_first_token_ms":180,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0133,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2457,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":497,"p95_latency_ms":608,"p99_latency_ms":652,"time_to_first_token_ms":183,"total_runtime_ms":553,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0114,"confidence_interval":0.9355,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9932,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":568,"p95_latency_ms":695,"p99_latency_ms":745,"time_to_first_token_ms":210,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.025,"confidence_interval":0.9488,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4567,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":476,"p95_latency_ms":582,"p99_latency_ms":624,"time_to_first_token_ms":190,"total_runtime_ms":529,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0167,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9439,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":504,"p95_latency_ms":616,"p99_latency_ms":661,"time_to_first_token_ms":186,"total_runtime_ms":560,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0181,"confidence_interval":0.9433,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2935,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":564,"p95_latency_ms":689,"p99_latency_ms":739,"time_to_first_token_ms":206,"total_runtime_ms":626,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0116,"confidence_interval":0.9323,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4672,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":515,"p95_latency_ms":630,"p99_latency_ms":675,"time_to_first_token_ms":199,"total_runtime_ms":572,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0101,"confidence_interval":0.9388,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4131,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":558,"p95_latency_ms":682,"p99_latency_ms":732,"time_to_first_token_ms":180,"total_runtime_ms":620,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0281,"confidence_interval":0.9361,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0998,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:7b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":556,"p95_latency_ms":680,"p99_latency_ms":729,"time_to_first_token_ms":196,"total_runtime_ms":618,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0116,"confidence_interval":0.9302,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1662,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1454,"p95_latency_ms":1777,"p99_latency_ms":1906,"time_to_first_token_ms":331,"total_runtime_ms":1615,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0137,"confidence_interval":0.9433,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7525,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1443,"p95_latency_ms":1764,"p99_latency_ms":1892,"time_to_first_token_ms":330,"total_runtime_ms":1603,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0102,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1198,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1404,"p95_latency_ms":1716,"p99_latency_ms":1841,"time_to_first_token_ms":313,"total_runtime_ms":1560,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0246,"confidence_interval":0.946,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8617,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1454,"p95_latency_ms":1777,"p99_latency_ms":1906,"time_to_first_token_ms":322,"total_runtime_ms":1615,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0289,"confidence_interval":0.9316,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3166,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1460,"p95_latency_ms":1785,"p99_latency_ms":1914,"time_to_first_token_ms":316,"total_runtime_ms":1622,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.9353,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2916,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1476,"p95_latency_ms":1804,"p99_latency_ms":1935,"time_to_first_token_ms":328,"total_runtime_ms":1640,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0252,"confidence_interval":0.9487,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1442,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1434,"p95_latency_ms":1752,"p99_latency_ms":1880,"time_to_first_token_ms":301,"total_runtime_ms":1593,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0112,"confidence_interval":0.9481,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.423,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1414,"p95_latency_ms":1729,"p99_latency_ms":1854,"time_to_first_token_ms":318,"total_runtime_ms":1571,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0152,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4939,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1450,"p95_latency_ms":1772,"p99_latency_ms":1901,"time_to_first_token_ms":329,"total_runtime_ms":1611,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0286,"confidence_interval":0.9307,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7592,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1427,"p95_latency_ms":1744,"p99_latency_ms":1871,"time_to_first_token_ms":318,"total_runtime_ms":1585,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0198,"confidence_interval":0.9371,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0211,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":761,"p95_latency_ms":930,"p99_latency_ms":998,"time_to_first_token_ms":267,"total_runtime_ms":845,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0231,"confidence_interval":0.9472,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8657,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":810,"p95_latency_ms":990,"p99_latency_ms":1062,"time_to_first_token_ms":260,"total_runtime_ms":900,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0213,"confidence_interval":0.9414,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9873,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":777,"p95_latency_ms":950,"p99_latency_ms":1019,"time_to_first_token_ms":260,"total_runtime_ms":863,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0279,"confidence_interval":0.9486,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1937,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":761,"p95_latency_ms":930,"p99_latency_ms":998,"time_to_first_token_ms":286,"total_runtime_ms":846,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0258,"confidence_interval":0.9314,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5353,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":802,"p95_latency_ms":981,"p99_latency_ms":1052,"time_to_first_token_ms":283,"total_runtime_ms":891,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0199,"confidence_interval":0.9336,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2744,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1010,"time_to_first_token_ms":282,"total_runtime_ms":856,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.02,"confidence_interval":0.9487,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7227,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":735,"p95_latency_ms":898,"p99_latency_ms":963,"time_to_first_token_ms":256,"total_runtime_ms":816,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0211,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7568,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":755,"p95_latency_ms":923,"p99_latency_ms":990,"time_to_first_token_ms":288,"total_runtime_ms":839,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0173,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1114,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":799,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":261,"total_runtime_ms":887,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.9385,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.958,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"summarize_contextual|high|qwen2:14b|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":804,"p95_latency_ms":982,"p99_latency_ms":1054,"time_to_first_token_ms":261,"total_runtime_ms":893,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0197,"confidence_interval":0.9411,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4828,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":985,"p95_latency_ms":1204,"p99_latency_ms":1291,"time_to_first_token_ms":208,"total_runtime_ms":1094,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0234,"confidence_interval":0.9428,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5571,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":997,"p95_latency_ms":1219,"p99_latency_ms":1307,"time_to_first_token_ms":221,"total_runtime_ms":1108,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3474,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":989,"p95_latency_ms":1209,"p99_latency_ms":1297,"time_to_first_token_ms":204,"total_runtime_ms":1099,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.017,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9313,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1006,"p95_latency_ms":1230,"p99_latency_ms":1319,"time_to_first_token_ms":219,"total_runtime_ms":1118,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0292,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0803,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1027,"p95_latency_ms":1255,"p99_latency_ms":1347,"time_to_first_token_ms":218,"total_runtime_ms":1141,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0249,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3876,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":937,"p95_latency_ms":1145,"p99_latency_ms":1229,"time_to_first_token_ms":222,"total_runtime_ms":1041,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0127,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5731,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":965,"p95_latency_ms":1179,"p99_latency_ms":1265,"time_to_first_token_ms":191,"total_runtime_ms":1072,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0243,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6263,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":955,"p95_latency_ms":1168,"p99_latency_ms":1253,"time_to_first_token_ms":220,"total_runtime_ms":1061,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0227,"confidence_interval":0.9438,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3975,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1006,"p95_latency_ms":1229,"p99_latency_ms":1318,"time_to_first_token_ms":206,"total_runtime_ms":1117,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0259,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1555,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1042,"p95_latency_ms":1274,"p99_latency_ms":1367,"time_to_first_token_ms":216,"total_runtime_ms":1158,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0149,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6777,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":568,"p95_latency_ms":695,"p99_latency_ms":745,"time_to_first_token_ms":177,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0204,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8278,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":647,"p95_latency_ms":790,"p99_latency_ms":848,"time_to_first_token_ms":175,"total_runtime_ms":719,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0202,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4979,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":563,"p95_latency_ms":689,"p99_latency_ms":739,"time_to_first_token_ms":144,"total_runtime_ms":626,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0238,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0578,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":665,"p95_latency_ms":813,"p99_latency_ms":872,"time_to_first_token_ms":167,"total_runtime_ms":739,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2113,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":146,"total_runtime_ms":663,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0225,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3415,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":755,"time_to_first_token_ms":162,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0119,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6863,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":580,"p95_latency_ms":708,"p99_latency_ms":760,"time_to_first_token_ms":171,"total_runtime_ms":644,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.014,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2431,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":785,"time_to_first_token_ms":158,"total_runtime_ms":666,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0235,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8462,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":619,"p95_latency_ms":757,"p99_latency_ms":812,"time_to_first_token_ms":152,"total_runtime_ms":688,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0163,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5043,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":639,"p95_latency_ms":781,"p99_latency_ms":838,"time_to_first_token_ms":171,"total_runtime_ms":710,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0269,"confidence_interval":0.95,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.641,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1292,"p95_latency_ms":1580,"p99_latency_ms":1695,"time_to_first_token_ms":243,"total_runtime_ms":1436,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0174,"confidence_interval":0.9348,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4684,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1305,"p95_latency_ms":1595,"p99_latency_ms":1711,"time_to_first_token_ms":267,"total_runtime_ms":1450,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0146,"confidence_interval":0.9367,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0946,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1277,"p95_latency_ms":1561,"p99_latency_ms":1675,"time_to_first_token_ms":256,"total_runtime_ms":1419,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0104,"confidence_interval":0.9376,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2303,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1352,"p95_latency_ms":1653,"p99_latency_ms":1773,"time_to_first_token_ms":264,"total_runtime_ms":1503,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0163,"confidence_interval":0.944,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7795,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1285,"p95_latency_ms":1571,"p99_latency_ms":1685,"time_to_first_token_ms":262,"total_runtime_ms":1428,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0198,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7141,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1326,"p95_latency_ms":1621,"p99_latency_ms":1739,"time_to_first_token_ms":239,"total_runtime_ms":1474,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0238,"confidence_interval":0.9466,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0781,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1301,"p95_latency_ms":1590,"p99_latency_ms":1706,"time_to_first_token_ms":250,"total_runtime_ms":1446,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0168,"confidence_interval":0.9469,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6655,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1365,"p95_latency_ms":1668,"p99_latency_ms":1790,"time_to_first_token_ms":243,"total_runtime_ms":1517,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0114,"confidence_interval":0.937,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0312,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1303,"p95_latency_ms":1592,"p99_latency_ms":1708,"time_to_first_token_ms":243,"total_runtime_ms":1448,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0254,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7023,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1293,"p95_latency_ms":1581,"p99_latency_ms":1696,"time_to_first_token_ms":240,"total_runtime_ms":1437,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0228,"confidence_interval":0.931,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6769,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":787,"p95_latency_ms":962,"p99_latency_ms":1032,"time_to_first_token_ms":213,"total_runtime_ms":874,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0283,"confidence_interval":0.9494,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8388,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":838,"p95_latency_ms":1024,"p99_latency_ms":1099,"time_to_first_token_ms":192,"total_runtime_ms":931,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0266,"confidence_interval":0.9493,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4947,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":772,"p95_latency_ms":944,"p99_latency_ms":1012,"time_to_first_token_ms":207,"total_runtime_ms":858,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0156,"confidence_interval":0.941,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5681,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":805,"p95_latency_ms":983,"p99_latency_ms":1055,"time_to_first_token_ms":189,"total_runtime_ms":894,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0164,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4721,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":788,"p95_latency_ms":963,"p99_latency_ms":1033,"time_to_first_token_ms":206,"total_runtime_ms":875,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0211,"confidence_interval":0.9391,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0002,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":858,"p95_latency_ms":1048,"p99_latency_ms":1125,"time_to_first_token_ms":200,"total_runtime_ms":953,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0252,"confidence_interval":0.9322,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5501,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":782,"p95_latency_ms":956,"p99_latency_ms":1025,"time_to_first_token_ms":217,"total_runtime_ms":869,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0216,"confidence_interval":0.9365,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0177,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":786,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":191,"total_runtime_ms":874,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0185,"confidence_interval":0.9418,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1846,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":821,"p95_latency_ms":1003,"p99_latency_ms":1076,"time_to_first_token_ms":200,"total_runtime_ms":912,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0225,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1101,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":844,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":190,"total_runtime_ms":937,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0144,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0962,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2529,"p95_latency_ms":3091,"p99_latency_ms":3316,"time_to_first_token_ms":438,"total_runtime_ms":2810,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.017,"confidence_interval":0.9459,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3767,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2633,"p95_latency_ms":3218,"p99_latency_ms":3452,"time_to_first_token_ms":417,"total_runtime_ms":2926,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0203,"confidence_interval":0.949,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9555,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2609,"p95_latency_ms":3189,"p99_latency_ms":3421,"time_to_first_token_ms":426,"total_runtime_ms":2899,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0241,"confidence_interval":0.9485,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3625,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2569,"p95_latency_ms":3140,"p99_latency_ms":3369,"time_to_first_token_ms":437,"total_runtime_ms":2855,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0113,"confidence_interval":0.9453,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3792,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2578,"p95_latency_ms":3151,"p99_latency_ms":3380,"time_to_first_token_ms":433,"total_runtime_ms":2865,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0155,"confidence_interval":0.9318,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1438,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2557,"p95_latency_ms":3126,"p99_latency_ms":3353,"time_to_first_token_ms":428,"total_runtime_ms":2841,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0236,"confidence_interval":0.9326,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5697,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2541,"p95_latency_ms":3105,"p99_latency_ms":3331,"time_to_first_token_ms":418,"total_runtime_ms":2823,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0269,"confidence_interval":0.9425,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5318,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2534,"p95_latency_ms":3097,"p99_latency_ms":3322,"time_to_first_token_ms":411,"total_runtime_ms":2815,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8707,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2572,"p95_latency_ms":3143,"p99_latency_ms":3372,"time_to_first_token_ms":437,"total_runtime_ms":2857,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0179,"confidence_interval":0.9464,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.795,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2552,"p95_latency_ms":3119,"p99_latency_ms":3346,"time_to_first_token_ms":423,"total_runtime_ms":2835,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0234,"confidence_interval":0.9431,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7855,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1590,"p95_latency_ms":1943,"p99_latency_ms":2084,"time_to_first_token_ms":387,"total_runtime_ms":1766,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0238,"confidence_interval":0.9428,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4059,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1586,"p95_latency_ms":1938,"p99_latency_ms":2079,"time_to_first_token_ms":366,"total_runtime_ms":1762,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0172,"confidence_interval":0.9421,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.481,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1519,"p95_latency_ms":1857,"p99_latency_ms":1992,"time_to_first_token_ms":361,"total_runtime_ms":1688,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0201,"confidence_interval":0.9481,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7289,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1526,"p95_latency_ms":1865,"p99_latency_ms":2001,"time_to_first_token_ms":373,"total_runtime_ms":1695,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0135,"confidence_interval":0.9422,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4437,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1621,"p95_latency_ms":1982,"p99_latency_ms":2126,"time_to_first_token_ms":388,"total_runtime_ms":1802,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0145,"confidence_interval":0.938,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6274,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1563,"p95_latency_ms":1911,"p99_latency_ms":2050,"time_to_first_token_ms":379,"total_runtime_ms":1737,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0178,"confidence_interval":0.9494,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4028,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1621,"p95_latency_ms":1981,"p99_latency_ms":2125,"time_to_first_token_ms":367,"total_runtime_ms":1801,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0249,"confidence_interval":0.9435,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1351,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1566,"p95_latency_ms":1914,"p99_latency_ms":2053,"time_to_first_token_ms":395,"total_runtime_ms":1740,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0284,"confidence_interval":0.9447,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1883,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1623,"p95_latency_ms":1984,"p99_latency_ms":2128,"time_to_first_token_ms":379,"total_runtime_ms":1803,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0114,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2503,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:7b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1617,"p95_latency_ms":1976,"p99_latency_ms":2119,"time_to_first_token_ms":362,"total_runtime_ms":1796,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0243,"confidence_interval":0.9434,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3564,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3823,"p95_latency_ms":4673,"p99_latency_ms":5013,"time_to_first_token_ms":574,"total_runtime_ms":4248,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0296,"confidence_interval":0.9428,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8701,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3821,"p95_latency_ms":4670,"p99_latency_ms":5010,"time_to_first_token_ms":605,"total_runtime_ms":4246,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.9352,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2402,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3765,"p95_latency_ms":4602,"p99_latency_ms":4937,"time_to_first_token_ms":592,"total_runtime_ms":4184,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0153,"confidence_interval":0.9411,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4357,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3765,"p95_latency_ms":4602,"p99_latency_ms":4936,"time_to_first_token_ms":599,"total_runtime_ms":4183,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0242,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7093,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3802,"p95_latency_ms":4647,"p99_latency_ms":4985,"time_to_first_token_ms":604,"total_runtime_ms":4224,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0171,"confidence_interval":0.9453,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5712,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3763,"p95_latency_ms":4600,"p99_latency_ms":4934,"time_to_first_token_ms":574,"total_runtime_ms":4182,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0136,"confidence_interval":0.9437,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4325,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3790,"p95_latency_ms":4633,"p99_latency_ms":4970,"time_to_first_token_ms":574,"total_runtime_ms":4212,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0153,"confidence_interval":0.9399,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9738,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3788,"p95_latency_ms":4630,"p99_latency_ms":4966,"time_to_first_token_ms":589,"total_runtime_ms":4209,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0237,"confidence_interval":0.9433,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0595,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3795,"p95_latency_ms":4638,"p99_latency_ms":4975,"time_to_first_token_ms":576,"total_runtime_ms":4216,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0167,"confidence_interval":0.9361,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3463,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3762,"p95_latency_ms":4598,"p99_latency_ms":4933,"time_to_first_token_ms":594,"total_runtime_ms":4180,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0282,"confidence_interval":0.9408,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.589,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2289,"p95_latency_ms":2798,"p99_latency_ms":3001,"time_to_first_token_ms":532,"total_runtime_ms":2543,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0215,"confidence_interval":0.9438,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9574,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2350,"p95_latency_ms":2872,"p99_latency_ms":3081,"time_to_first_token_ms":539,"total_runtime_ms":2611,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0247,"confidence_interval":0.9383,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.198,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2283,"p95_latency_ms":2790,"p99_latency_ms":2993,"time_to_first_token_ms":556,"total_runtime_ms":2537,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0139,"confidence_interval":0.9362,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3959,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2294,"p95_latency_ms":2804,"p99_latency_ms":3008,"time_to_first_token_ms":545,"total_runtime_ms":2549,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0148,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6622,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2305,"p95_latency_ms":2818,"p99_latency_ms":3023,"time_to_first_token_ms":545,"total_runtime_ms":2562,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0275,"confidence_interval":0.9398,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0084,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2311,"p95_latency_ms":2825,"p99_latency_ms":3030,"time_to_first_token_ms":535,"total_runtime_ms":2568,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0133,"confidence_interval":0.9314,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5287,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2353,"p95_latency_ms":2876,"p99_latency_ms":3085,"time_to_first_token_ms":553,"total_runtime_ms":2615,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0182,"confidence_interval":0.9467,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5131,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2282,"p95_latency_ms":2789,"p99_latency_ms":2991,"time_to_first_token_ms":543,"total_runtime_ms":2535,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0261,"confidence_interval":0.9401,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0205,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2274,"p95_latency_ms":2780,"p99_latency_ms":2982,"time_to_first_token_ms":562,"total_runtime_ms":2527,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0205,"confidence_interval":0.9397,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4555,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|low|qwen2:14b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2282,"p95_latency_ms":2789,"p99_latency_ms":2992,"time_to_first_token_ms":545,"total_runtime_ms":2536,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4698,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":727,"p95_latency_ms":889,"p99_latency_ms":954,"time_to_first_token_ms":160,"total_runtime_ms":808,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0203,"confidence_interval":0.9397,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7334,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":785,"p95_latency_ms":960,"p99_latency_ms":1030,"time_to_first_token_ms":189,"total_runtime_ms":873,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0262,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8161,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":760,"p95_latency_ms":929,"p99_latency_ms":996,"time_to_first_token_ms":165,"total_runtime_ms":844,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.029,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2623,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":790,"p95_latency_ms":966,"p99_latency_ms":1036,"time_to_first_token_ms":187,"total_runtime_ms":878,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.028,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7495,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":743,"p95_latency_ms":909,"p99_latency_ms":975,"time_to_first_token_ms":188,"total_runtime_ms":826,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0217,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6147,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":782,"p95_latency_ms":956,"p99_latency_ms":1025,"time_to_first_token_ms":186,"total_runtime_ms":869,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0176,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5938,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":807,"p95_latency_ms":987,"p99_latency_ms":1058,"time_to_first_token_ms":163,"total_runtime_ms":897,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0109,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5094,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":736,"p95_latency_ms":900,"p99_latency_ms":965,"time_to_first_token_ms":187,"total_runtime_ms":818,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0283,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.834,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":795,"p95_latency_ms":971,"p99_latency_ms":1042,"time_to_first_token_ms":165,"total_runtime_ms":883,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0233,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9062,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":766,"p95_latency_ms":936,"p99_latency_ms":1004,"time_to_first_token_ms":168,"total_runtime_ms":851,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0103,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9584,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":511,"p95_latency_ms":624,"p99_latency_ms":670,"time_to_first_token_ms":130,"total_runtime_ms":567,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0234,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.757,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":515,"p95_latency_ms":629,"p99_latency_ms":675,"time_to_first_token_ms":128,"total_runtime_ms":572,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0136,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8216,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":555,"time_to_first_token_ms":128,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0103,"confidence_interval":0.9305,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8496,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":449,"p95_latency_ms":549,"p99_latency_ms":589,"time_to_first_token_ms":135,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0172,"confidence_interval":0.9428,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6073,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":124,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0143,"confidence_interval":0.9465,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6762,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":457,"p95_latency_ms":559,"p99_latency_ms":600,"time_to_first_token_ms":143,"total_runtime_ms":508,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0145,"confidence_interval":0.9388,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2191,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":520,"p95_latency_ms":635,"p99_latency_ms":682,"time_to_first_token_ms":119,"total_runtime_ms":578,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9565,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":140,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0192,"confidence_interval":0.95,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7042,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":486,"p95_latency_ms":594,"p99_latency_ms":638,"time_to_first_token_ms":145,"total_runtime_ms":540,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0103,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1905,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":516,"p95_latency_ms":631,"p99_latency_ms":677,"time_to_first_token_ms":146,"total_runtime_ms":574,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0223,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9213,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":958,"p95_latency_ms":1171,"p99_latency_ms":1256,"time_to_first_token_ms":196,"total_runtime_ms":1064,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0137,"confidence_interval":0.9434,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7073,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1019,"p95_latency_ms":1245,"p99_latency_ms":1336,"time_to_first_token_ms":202,"total_runtime_ms":1132,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0213,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7909,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":983,"p95_latency_ms":1202,"p99_latency_ms":1289,"time_to_first_token_ms":220,"total_runtime_ms":1092,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0111,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3724,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":989,"p95_latency_ms":1209,"p99_latency_ms":1297,"time_to_first_token_ms":203,"total_runtime_ms":1099,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0135,"confidence_interval":0.9362,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8527,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1001,"p95_latency_ms":1223,"p99_latency_ms":1312,"time_to_first_token_ms":193,"total_runtime_ms":1112,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0151,"confidence_interval":0.9429,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4548,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":962,"p95_latency_ms":1175,"p99_latency_ms":1261,"time_to_first_token_ms":199,"total_runtime_ms":1068,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0226,"confidence_interval":0.9431,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8035,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":992,"p95_latency_ms":1213,"p99_latency_ms":1301,"time_to_first_token_ms":206,"total_runtime_ms":1103,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0211,"confidence_interval":0.9395,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9283,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":962,"p95_latency_ms":1175,"p99_latency_ms":1261,"time_to_first_token_ms":192,"total_runtime_ms":1068,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0146,"confidence_interval":0.9387,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3831,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":966,"p95_latency_ms":1181,"p99_latency_ms":1267,"time_to_first_token_ms":202,"total_runtime_ms":1073,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0216,"confidence_interval":0.9434,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.38,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1028,"p95_latency_ms":1257,"p99_latency_ms":1348,"time_to_first_token_ms":217,"total_runtime_ms":1142,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9402,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0732,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":617,"p95_latency_ms":754,"p99_latency_ms":808,"time_to_first_token_ms":158,"total_runtime_ms":685,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0225,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0903,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":164,"total_runtime_ms":636,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0173,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5188,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":673,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":165,"total_runtime_ms":748,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0283,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9005,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":617,"p95_latency_ms":754,"p99_latency_ms":809,"time_to_first_token_ms":152,"total_runtime_ms":686,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0149,"confidence_interval":0.934,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6528,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":674,"p95_latency_ms":823,"p99_latency_ms":883,"time_to_first_token_ms":175,"total_runtime_ms":749,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0118,"confidence_interval":0.9492,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6441,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":570,"p95_latency_ms":697,"p99_latency_ms":748,"time_to_first_token_ms":160,"total_runtime_ms":634,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.012,"confidence_interval":0.9471,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1184,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":780,"time_to_first_token_ms":164,"total_runtime_ms":661,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0179,"confidence_interval":0.9335,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1587,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":637,"p95_latency_ms":778,"p99_latency_ms":835,"time_to_first_token_ms":158,"total_runtime_ms":707,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0289,"confidence_interval":0.9376,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0542,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":613,"p95_latency_ms":749,"p99_latency_ms":803,"time_to_first_token_ms":148,"total_runtime_ms":681,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0117,"confidence_interval":0.9365,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1139,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":630,"p95_latency_ms":769,"p99_latency_ms":825,"time_to_first_token_ms":161,"total_runtime_ms":699,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.023,"confidence_interval":0.938,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4816,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1938,"p95_latency_ms":2368,"p99_latency_ms":2540,"time_to_first_token_ms":349,"total_runtime_ms":2153,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0174,"confidence_interval":0.9343,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4026,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1949,"p95_latency_ms":2382,"p99_latency_ms":2555,"time_to_first_token_ms":319,"total_runtime_ms":2165,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0232,"confidence_interval":0.9351,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2457,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1948,"p95_latency_ms":2381,"p99_latency_ms":2554,"time_to_first_token_ms":349,"total_runtime_ms":2165,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.016,"confidence_interval":0.9474,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9497,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1918,"p95_latency_ms":2344,"p99_latency_ms":2515,"time_to_first_token_ms":322,"total_runtime_ms":2131,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0243,"confidence_interval":0.9452,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2851,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1932,"p95_latency_ms":2361,"p99_latency_ms":2533,"time_to_first_token_ms":320,"total_runtime_ms":2146,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0115,"confidence_interval":0.9362,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9021,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1995,"p95_latency_ms":2438,"p99_latency_ms":2616,"time_to_first_token_ms":340,"total_runtime_ms":2217,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.01,"confidence_interval":0.9497,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.98,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1940,"p95_latency_ms":2371,"p99_latency_ms":2544,"time_to_first_token_ms":341,"total_runtime_ms":2156,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0193,"confidence_interval":0.9495,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3158,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1904,"p95_latency_ms":2327,"p99_latency_ms":2496,"time_to_first_token_ms":320,"total_runtime_ms":2115,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0208,"confidence_interval":0.9389,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7466,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2000,"p95_latency_ms":2445,"p99_latency_ms":2622,"time_to_first_token_ms":337,"total_runtime_ms":2222,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0265,"confidence_interval":0.9479,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7968,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1952,"p95_latency_ms":2385,"p99_latency_ms":2559,"time_to_first_token_ms":326,"total_runtime_ms":2169,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0231,"confidence_interval":0.9412,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7365,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1212,"p95_latency_ms":1482,"p99_latency_ms":1590,"time_to_first_token_ms":284,"total_runtime_ms":1347,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0155,"confidence_interval":0.9323,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4756,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1157,"p95_latency_ms":1415,"p99_latency_ms":1518,"time_to_first_token_ms":276,"total_runtime_ms":1286,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0251,"confidence_interval":0.9474,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8838,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1157,"p95_latency_ms":1414,"p99_latency_ms":1517,"time_to_first_token_ms":306,"total_runtime_ms":1285,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.026,"confidence_interval":0.9497,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0708,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1193,"p95_latency_ms":1458,"p99_latency_ms":1564,"time_to_first_token_ms":291,"total_runtime_ms":1325,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0275,"confidence_interval":0.9332,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8447,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1140,"p95_latency_ms":1394,"p99_latency_ms":1495,"time_to_first_token_ms":276,"total_runtime_ms":1267,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0159,"confidence_interval":0.9381,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3553,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1214,"p95_latency_ms":1484,"p99_latency_ms":1592,"time_to_first_token_ms":284,"total_runtime_ms":1349,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0207,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1166,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1215,"p95_latency_ms":1485,"p99_latency_ms":1593,"time_to_first_token_ms":303,"total_runtime_ms":1350,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0185,"confidence_interval":0.9399,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9946,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1164,"p95_latency_ms":1423,"p99_latency_ms":1526,"time_to_first_token_ms":282,"total_runtime_ms":1293,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0133,"confidence_interval":0.946,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1981,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1215,"p95_latency_ms":1485,"p99_latency_ms":1593,"time_to_first_token_ms":296,"total_runtime_ms":1350,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0119,"confidence_interval":0.9467,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4634,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1170,"p95_latency_ms":1430,"p99_latency_ms":1534,"time_to_first_token_ms":279,"total_runtime_ms":1300,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0101,"confidence_interval":0.9361,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9225,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2875,"p95_latency_ms":3513,"p99_latency_ms":3769,"time_to_first_token_ms":457,"total_runtime_ms":3194,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0209,"confidence_interval":0.9443,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1215,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2848,"p95_latency_ms":3481,"p99_latency_ms":3734,"time_to_first_token_ms":449,"total_runtime_ms":3164,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9386,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9359,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2905,"p95_latency_ms":3550,"p99_latency_ms":3808,"time_to_first_token_ms":459,"total_runtime_ms":3228,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0283,"confidence_interval":0.9485,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9334,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2822,"p95_latency_ms":3449,"p99_latency_ms":3700,"time_to_first_token_ms":461,"total_runtime_ms":3136,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0157,"confidence_interval":0.9484,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7007,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2818,"p95_latency_ms":3444,"p99_latency_ms":3694,"time_to_first_token_ms":467,"total_runtime_ms":3131,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0286,"confidence_interval":0.9419,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7379,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2834,"p95_latency_ms":3464,"p99_latency_ms":3716,"time_to_first_token_ms":461,"total_runtime_ms":3149,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0249,"confidence_interval":0.9333,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.658,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2895,"p95_latency_ms":3538,"p99_latency_ms":3795,"time_to_first_token_ms":462,"total_runtime_ms":3216,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0136,"confidence_interval":0.9319,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1495,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2829,"p95_latency_ms":3458,"p99_latency_ms":3709,"time_to_first_token_ms":475,"total_runtime_ms":3143,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0103,"confidence_interval":0.9331,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4849,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2911,"p95_latency_ms":3558,"p99_latency_ms":3817,"time_to_first_token_ms":467,"total_runtime_ms":3235,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0128,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7998,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2847,"p95_latency_ms":3480,"p99_latency_ms":3733,"time_to_first_token_ms":458,"total_runtime_ms":3163,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0191,"confidence_interval":0.9359,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.968,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1686,"p95_latency_ms":2060,"p99_latency_ms":2210,"time_to_first_token_ms":403,"total_runtime_ms":1873,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.027,"confidence_interval":0.9408,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3137,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1782,"p95_latency_ms":2178,"p99_latency_ms":2337,"time_to_first_token_ms":412,"total_runtime_ms":1980,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0126,"confidence_interval":0.9487,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3293,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1728,"p95_latency_ms":2111,"p99_latency_ms":2265,"time_to_first_token_ms":423,"total_runtime_ms":1919,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0105,"confidence_interval":0.939,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9485,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1707,"p95_latency_ms":2087,"p99_latency_ms":2238,"time_to_first_token_ms":430,"total_runtime_ms":1897,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.023,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7686,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1728,"p95_latency_ms":2112,"p99_latency_ms":2266,"time_to_first_token_ms":401,"total_runtime_ms":1920,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0245,"confidence_interval":0.9343,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5141,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1755,"p95_latency_ms":2145,"p99_latency_ms":2301,"time_to_first_token_ms":425,"total_runtime_ms":1950,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9457,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9271,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1731,"p95_latency_ms":2115,"p99_latency_ms":2269,"time_to_first_token_ms":412,"total_runtime_ms":1923,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0299,"confidence_interval":0.9423,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3777,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1701,"p95_latency_ms":2079,"p99_latency_ms":2230,"time_to_first_token_ms":424,"total_runtime_ms":1890,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0145,"confidence_interval":0.9332,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0823,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1699,"p95_latency_ms":2076,"p99_latency_ms":2227,"time_to_first_token_ms":410,"total_runtime_ms":1887,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0231,"confidence_interval":0.937,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1299,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1718,"p95_latency_ms":2100,"p99_latency_ms":2253,"time_to_first_token_ms":414,"total_runtime_ms":1909,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.012,"confidence_interval":0.9461,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7419,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":161,"total_runtime_ms":586,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0254,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1245,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":558,"p95_latency_ms":682,"p99_latency_ms":732,"time_to_first_token_ms":144,"total_runtime_ms":620,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0163,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7458,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":538,"p95_latency_ms":658,"p99_latency_ms":706,"time_to_first_token_ms":133,"total_runtime_ms":598,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0185,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4639,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":140,"total_runtime_ms":591,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0165,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2164,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":597,"p95_latency_ms":729,"p99_latency_ms":782,"time_to_first_token_ms":132,"total_runtime_ms":663,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0199,"confidence_interval":0.9495,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0258,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":545,"p95_latency_ms":666,"p99_latency_ms":715,"time_to_first_token_ms":157,"total_runtime_ms":606,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0178,"confidence_interval":0.944,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3235,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":564,"p95_latency_ms":690,"p99_latency_ms":740,"time_to_first_token_ms":155,"total_runtime_ms":627,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0196,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1666,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":629,"p95_latency_ms":769,"p99_latency_ms":825,"time_to_first_token_ms":148,"total_runtime_ms":699,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0205,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3231,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":145,"total_runtime_ms":664,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0112,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4922,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":633,"p95_latency_ms":774,"p99_latency_ms":830,"time_to_first_token_ms":144,"total_runtime_ms":703,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0174,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3056,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":368,"p95_latency_ms":449,"p99_latency_ms":482,"time_to_first_token_ms":117,"total_runtime_ms":408,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.027,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2816,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":422,"p95_latency_ms":516,"p99_latency_ms":554,"time_to_first_token_ms":90,"total_runtime_ms":469,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1986,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":415,"p95_latency_ms":507,"p99_latency_ms":544,"time_to_first_token_ms":97,"total_runtime_ms":461,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0221,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9266,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":120,"total_runtime_ms":453,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0118,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0443,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":350,"p95_latency_ms":428,"p99_latency_ms":459,"time_to_first_token_ms":111,"total_runtime_ms":389,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0163,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0118,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":414,"p95_latency_ms":506,"p99_latency_ms":542,"time_to_first_token_ms":92,"total_runtime_ms":460,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0245,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8558,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":368,"p95_latency_ms":449,"p99_latency_ms":482,"time_to_first_token_ms":94,"total_runtime_ms":408,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0103,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7233,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":330,"p95_latency_ms":403,"p99_latency_ms":433,"time_to_first_token_ms":110,"total_runtime_ms":367,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.017,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1046,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":383,"p95_latency_ms":468,"p99_latency_ms":502,"time_to_first_token_ms":98,"total_runtime_ms":426,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0267,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.317,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":375,"p95_latency_ms":459,"p99_latency_ms":492,"time_to_first_token_ms":98,"total_runtime_ms":417,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9502,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":795,"p95_latency_ms":972,"p99_latency_ms":1043,"time_to_first_token_ms":190,"total_runtime_ms":884,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.026,"confidence_interval":0.9331,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5633,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":818,"p95_latency_ms":1000,"p99_latency_ms":1073,"time_to_first_token_ms":160,"total_runtime_ms":909,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0223,"confidence_interval":0.9305,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.308,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":779,"p95_latency_ms":952,"p99_latency_ms":1021,"time_to_first_token_ms":164,"total_runtime_ms":865,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0204,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3098,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":759,"p95_latency_ms":928,"p99_latency_ms":996,"time_to_first_token_ms":168,"total_runtime_ms":844,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0163,"confidence_interval":0.9365,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7117,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":795,"p95_latency_ms":972,"p99_latency_ms":1042,"time_to_first_token_ms":174,"total_runtime_ms":883,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0136,"confidence_interval":0.9348,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.687,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":747,"p95_latency_ms":913,"p99_latency_ms":979,"time_to_first_token_ms":166,"total_runtime_ms":830,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0238,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5648,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":746,"p95_latency_ms":912,"p99_latency_ms":978,"time_to_first_token_ms":165,"total_runtime_ms":829,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0242,"confidence_interval":0.9421,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6173,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":715,"p95_latency_ms":874,"p99_latency_ms":938,"time_to_first_token_ms":159,"total_runtime_ms":795,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0218,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3624,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":948,"time_to_first_token_ms":180,"total_runtime_ms":803,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0117,"confidence_interval":0.9343,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3149,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":745,"p95_latency_ms":910,"p99_latency_ms":977,"time_to_first_token_ms":160,"total_runtime_ms":828,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0253,"confidence_interval":0.9353,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4038,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":669,"time_to_first_token_ms":135,"total_runtime_ms":567,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0241,"confidence_interval":0.9443,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4249,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":672,"time_to_first_token_ms":119,"total_runtime_ms":569,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0168,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2678,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":645,"p99_latency_ms":691,"time_to_first_token_ms":140,"total_runtime_ms":586,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.016,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.252,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":520,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":138,"total_runtime_ms":578,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0212,"confidence_interval":0.9357,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9408,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":519,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":142,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0127,"confidence_interval":0.9405,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5004,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":504,"p95_latency_ms":616,"p99_latency_ms":661,"time_to_first_token_ms":133,"total_runtime_ms":560,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0275,"confidence_interval":0.9318,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1636,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":140,"total_runtime_ms":574,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.025,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1215,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":139,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0135,"confidence_interval":0.9364,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5394,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":440,"p95_latency_ms":537,"p99_latency_ms":577,"time_to_first_token_ms":124,"total_runtime_ms":489,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0142,"confidence_interval":0.9378,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.405,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":436,"p95_latency_ms":533,"p99_latency_ms":572,"time_to_first_token_ms":145,"total_runtime_ms":485,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0228,"confidence_interval":0.9446,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6375,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1474,"p95_latency_ms":1801,"p99_latency_ms":1932,"time_to_first_token_ms":264,"total_runtime_ms":1637,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0166,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7943,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1498,"p95_latency_ms":1830,"p99_latency_ms":1964,"time_to_first_token_ms":272,"total_runtime_ms":1664,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.02,"confidence_interval":0.9477,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.184,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1479,"p95_latency_ms":1807,"p99_latency_ms":1939,"time_to_first_token_ms":285,"total_runtime_ms":1643,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0163,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.643,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1497,"p95_latency_ms":1830,"p99_latency_ms":1963,"time_to_first_token_ms":263,"total_runtime_ms":1663,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.021,"confidence_interval":0.9348,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8671,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1467,"p95_latency_ms":1793,"p99_latency_ms":1923,"time_to_first_token_ms":267,"total_runtime_ms":1630,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0196,"confidence_interval":0.9437,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1549,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1485,"p95_latency_ms":1815,"p99_latency_ms":1947,"time_to_first_token_ms":266,"total_runtime_ms":1650,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0117,"confidence_interval":0.93,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7458,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1454,"p95_latency_ms":1777,"p99_latency_ms":1906,"time_to_first_token_ms":268,"total_runtime_ms":1616,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0168,"confidence_interval":0.9495,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8805,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1430,"p95_latency_ms":1747,"p99_latency_ms":1874,"time_to_first_token_ms":288,"total_runtime_ms":1588,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0105,"confidence_interval":0.9434,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2622,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1485,"p95_latency_ms":1815,"p99_latency_ms":1947,"time_to_first_token_ms":282,"total_runtime_ms":1650,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0184,"confidence_interval":0.9487,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7708,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1439,"p95_latency_ms":1759,"p99_latency_ms":1887,"time_to_first_token_ms":255,"total_runtime_ms":1599,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0158,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8727,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":879,"p95_latency_ms":1075,"p99_latency_ms":1153,"time_to_first_token_ms":228,"total_runtime_ms":977,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0182,"confidence_interval":0.9368,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4129,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":926,"p95_latency_ms":1132,"p99_latency_ms":1215,"time_to_first_token_ms":228,"total_runtime_ms":1029,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0115,"confidence_interval":0.9447,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.398,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":928,"p95_latency_ms":1134,"p99_latency_ms":1217,"time_to_first_token_ms":239,"total_runtime_ms":1031,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0145,"confidence_interval":0.9475,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0575,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":916,"p95_latency_ms":1120,"p99_latency_ms":1201,"time_to_first_token_ms":241,"total_runtime_ms":1018,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0247,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8407,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":953,"p95_latency_ms":1165,"p99_latency_ms":1249,"time_to_first_token_ms":237,"total_runtime_ms":1059,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0154,"confidence_interval":0.9328,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3919,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":870,"p95_latency_ms":1063,"p99_latency_ms":1140,"time_to_first_token_ms":221,"total_runtime_ms":966,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0292,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3306,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":925,"p95_latency_ms":1131,"p99_latency_ms":1213,"time_to_first_token_ms":240,"total_runtime_ms":1028,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0164,"confidence_interval":0.9367,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3974,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":883,"p95_latency_ms":1079,"p99_latency_ms":1157,"time_to_first_token_ms":229,"total_runtime_ms":981,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0282,"confidence_interval":0.9464,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4617,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":946,"p95_latency_ms":1156,"p99_latency_ms":1240,"time_to_first_token_ms":226,"total_runtime_ms":1051,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0251,"confidence_interval":0.9439,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4777,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":870,"p95_latency_ms":1063,"p99_latency_ms":1140,"time_to_first_token_ms":242,"total_runtime_ms":966,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0206,"confidence_interval":0.9361,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1366,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2185,"p95_latency_ms":2670,"p99_latency_ms":2865,"time_to_first_token_ms":365,"total_runtime_ms":2428,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0262,"confidence_interval":0.9436,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8859,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2213,"p95_latency_ms":2705,"p99_latency_ms":2901,"time_to_first_token_ms":377,"total_runtime_ms":2459,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.011,"confidence_interval":0.9473,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4608,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2174,"p95_latency_ms":2657,"p99_latency_ms":2850,"time_to_first_token_ms":383,"total_runtime_ms":2416,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0126,"confidence_interval":0.9489,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6349,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2170,"p95_latency_ms":2652,"p99_latency_ms":2845,"time_to_first_token_ms":373,"total_runtime_ms":2411,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0152,"confidence_interval":0.9422,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0307,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2189,"p95_latency_ms":2676,"p99_latency_ms":2870,"time_to_first_token_ms":356,"total_runtime_ms":2432,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0265,"confidence_interval":0.9472,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3275,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2203,"p95_latency_ms":2693,"p99_latency_ms":2889,"time_to_first_token_ms":356,"total_runtime_ms":2448,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0139,"confidence_interval":0.9332,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6253,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2144,"p95_latency_ms":2620,"p99_latency_ms":2811,"time_to_first_token_ms":369,"total_runtime_ms":2382,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0175,"confidence_interval":0.9353,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5084,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2146,"p95_latency_ms":2623,"p99_latency_ms":2814,"time_to_first_token_ms":372,"total_runtime_ms":2385,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0106,"confidence_interval":0.9485,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3822,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2188,"p95_latency_ms":2674,"p99_latency_ms":2868,"time_to_first_token_ms":373,"total_runtime_ms":2431,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0175,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8524,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2140,"p95_latency_ms":2615,"p99_latency_ms":2806,"time_to_first_token_ms":369,"total_runtime_ms":2378,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0236,"confidence_interval":0.9446,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1402,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1311,"p95_latency_ms":1602,"p99_latency_ms":1719,"time_to_first_token_ms":314,"total_runtime_ms":1457,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0169,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3862,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1343,"p95_latency_ms":1641,"p99_latency_ms":1760,"time_to_first_token_ms":332,"total_runtime_ms":1492,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0141,"confidence_interval":0.9375,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0001,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1329,"p95_latency_ms":1624,"p99_latency_ms":1743,"time_to_first_token_ms":310,"total_runtime_ms":1477,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0124,"confidence_interval":0.9371,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3673,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1319,"p95_latency_ms":1612,"p99_latency_ms":1729,"time_to_first_token_ms":319,"total_runtime_ms":1465,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0293,"confidence_interval":0.9346,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1663,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1299,"p95_latency_ms":1588,"p99_latency_ms":1703,"time_to_first_token_ms":319,"total_runtime_ms":1443,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0179,"confidence_interval":0.9497,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1026,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1329,"p95_latency_ms":1625,"p99_latency_ms":1743,"time_to_first_token_ms":325,"total_runtime_ms":1477,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0173,"confidence_interval":0.9496,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4876,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1362,"p95_latency_ms":1664,"p99_latency_ms":1785,"time_to_first_token_ms":315,"total_runtime_ms":1513,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0272,"confidence_interval":0.9349,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1938,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1285,"p95_latency_ms":1571,"p99_latency_ms":1685,"time_to_first_token_ms":330,"total_runtime_ms":1428,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0293,"confidence_interval":0.948,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3322,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1277,"p95_latency_ms":1560,"p99_latency_ms":1674,"time_to_first_token_ms":314,"total_runtime_ms":1418,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0183,"confidence_interval":0.948,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5571,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1289,"p95_latency_ms":1576,"p99_latency_ms":1690,"time_to_first_token_ms":305,"total_runtime_ms":1433,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.935,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8951,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":472,"p95_latency_ms":576,"p99_latency_ms":618,"time_to_first_token_ms":140,"total_runtime_ms":524,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0222,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3707,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":445,"p95_latency_ms":543,"p99_latency_ms":583,"time_to_first_token_ms":139,"total_runtime_ms":494,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0282,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.464,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":546,"p95_latency_ms":667,"p99_latency_ms":716,"time_to_first_token_ms":128,"total_runtime_ms":607,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0198,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1808,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":141,"total_runtime_ms":536,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0196,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8617,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":473,"p95_latency_ms":578,"p99_latency_ms":620,"time_to_first_token_ms":141,"total_runtime_ms":525,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0283,"confidence_interval":0.9442,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7581,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":631,"time_to_first_token_ms":137,"total_runtime_ms":535,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0115,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8879,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":516,"p95_latency_ms":631,"p99_latency_ms":677,"time_to_first_token_ms":153,"total_runtime_ms":574,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0107,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7993,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":441,"p95_latency_ms":540,"p99_latency_ms":579,"time_to_first_token_ms":122,"total_runtime_ms":491,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0271,"confidence_interval":0.9307,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6778,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":546,"p95_latency_ms":667,"p99_latency_ms":716,"time_to_first_token_ms":128,"total_runtime_ms":607,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0105,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4525,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":511,"p95_latency_ms":624,"p99_latency_ms":669,"time_to_first_token_ms":121,"total_runtime_ms":567,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0104,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7497,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":330,"p95_latency_ms":403,"p99_latency_ms":433,"time_to_first_token_ms":103,"total_runtime_ms":367,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0141,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1293,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":305,"p95_latency_ms":373,"p99_latency_ms":400,"time_to_first_token_ms":108,"total_runtime_ms":339,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0219,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0861,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":437,"p99_latency_ms":469,"time_to_first_token_ms":82,"total_runtime_ms":398,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0151,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0318,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":338,"p95_latency_ms":413,"p99_latency_ms":443,"time_to_first_token_ms":92,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.029,"confidence_interval":0.9423,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3504,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":425,"time_to_first_token_ms":78,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0132,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8005,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":318,"p95_latency_ms":389,"p99_latency_ms":418,"time_to_first_token_ms":109,"total_runtime_ms":354,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.012,"confidence_interval":0.944,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7769,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":303,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":107,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0257,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3877,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":271,"p95_latency_ms":331,"p99_latency_ms":355,"time_to_first_token_ms":86,"total_runtime_ms":301,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0287,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9626,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":320,"p95_latency_ms":391,"p99_latency_ms":419,"time_to_first_token_ms":94,"total_runtime_ms":355,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0287,"confidence_interval":0.938,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.659,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":318,"p95_latency_ms":388,"p99_latency_ms":417,"time_to_first_token_ms":79,"total_runtime_ms":353,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0138,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3639,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":622,"p95_latency_ms":760,"p99_latency_ms":816,"time_to_first_token_ms":166,"total_runtime_ms":691,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0161,"confidence_interval":0.9399,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3891,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":174,"total_runtime_ms":764,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0253,"confidence_interval":0.9341,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4485,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":652,"p95_latency_ms":797,"p99_latency_ms":855,"time_to_first_token_ms":158,"total_runtime_ms":725,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0121,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5106,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":594,"p95_latency_ms":726,"p99_latency_ms":779,"time_to_first_token_ms":151,"total_runtime_ms":660,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0119,"confidence_interval":0.9358,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1568,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":623,"p95_latency_ms":761,"p99_latency_ms":817,"time_to_first_token_ms":149,"total_runtime_ms":692,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.01,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6465,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":655,"p95_latency_ms":801,"p99_latency_ms":859,"time_to_first_token_ms":144,"total_runtime_ms":728,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0115,"confidence_interval":0.9394,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.456,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":637,"p95_latency_ms":779,"p99_latency_ms":835,"time_to_first_token_ms":145,"total_runtime_ms":708,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0258,"confidence_interval":0.9459,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0612,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":656,"p95_latency_ms":802,"p99_latency_ms":860,"time_to_first_token_ms":141,"total_runtime_ms":729,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0117,"confidence_interval":0.9424,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7905,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":609,"p95_latency_ms":744,"p99_latency_ms":798,"time_to_first_token_ms":157,"total_runtime_ms":676,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0294,"confidence_interval":0.937,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9098,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":695,"p95_latency_ms":849,"p99_latency_ms":911,"time_to_first_token_ms":171,"total_runtime_ms":772,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0221,"confidence_interval":0.9405,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8939,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":455,"p95_latency_ms":556,"p99_latency_ms":596,"time_to_first_token_ms":108,"total_runtime_ms":505,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.024,"confidence_interval":0.9307,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5322,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":399,"p95_latency_ms":488,"p99_latency_ms":524,"time_to_first_token_ms":110,"total_runtime_ms":444,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0127,"confidence_interval":0.9405,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3469,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":357,"p95_latency_ms":436,"p99_latency_ms":468,"time_to_first_token_ms":118,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.019,"confidence_interval":0.9469,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5949,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":479,"time_to_first_token_ms":104,"total_runtime_ms":406,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0272,"confidence_interval":0.9411,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1184,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":417,"p95_latency_ms":509,"p99_latency_ms":546,"time_to_first_token_ms":120,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0229,"confidence_interval":0.9306,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7232,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":450,"p95_latency_ms":550,"p99_latency_ms":590,"time_to_first_token_ms":122,"total_runtime_ms":500,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0272,"confidence_interval":0.9387,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0357,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":507,"time_to_first_token_ms":118,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0104,"confidence_interval":0.9433,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4528,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":371,"p95_latency_ms":454,"p99_latency_ms":487,"time_to_first_token_ms":117,"total_runtime_ms":412,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0187,"confidence_interval":0.9416,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6054,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":400,"p95_latency_ms":489,"p99_latency_ms":525,"time_to_first_token_ms":130,"total_runtime_ms":445,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0181,"confidence_interval":0.936,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8266,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":407,"p95_latency_ms":497,"p99_latency_ms":534,"time_to_first_token_ms":130,"total_runtime_ms":452,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.014,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4285,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1198,"p95_latency_ms":1464,"p99_latency_ms":1571,"time_to_first_token_ms":228,"total_runtime_ms":1331,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0284,"confidence_interval":0.9444,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9342,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1276,"p95_latency_ms":1560,"p99_latency_ms":1673,"time_to_first_token_ms":249,"total_runtime_ms":1418,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0185,"confidence_interval":0.94,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2006,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1253,"p95_latency_ms":1531,"p99_latency_ms":1642,"time_to_first_token_ms":241,"total_runtime_ms":1392,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0296,"confidence_interval":0.9493,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8743,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1270,"p95_latency_ms":1552,"p99_latency_ms":1665,"time_to_first_token_ms":239,"total_runtime_ms":1411,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0265,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0457,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1290,"p95_latency_ms":1577,"p99_latency_ms":1691,"time_to_first_token_ms":237,"total_runtime_ms":1433,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0135,"confidence_interval":0.9351,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0806,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1289,"p95_latency_ms":1576,"p99_latency_ms":1690,"time_to_first_token_ms":243,"total_runtime_ms":1432,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0229,"confidence_interval":0.9451,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6697,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1270,"p95_latency_ms":1553,"p99_latency_ms":1666,"time_to_first_token_ms":242,"total_runtime_ms":1412,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0268,"confidence_interval":0.9395,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.97,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1211,"p95_latency_ms":1480,"p99_latency_ms":1588,"time_to_first_token_ms":256,"total_runtime_ms":1345,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0128,"confidence_interval":0.9364,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8092,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1219,"p95_latency_ms":1490,"p99_latency_ms":1599,"time_to_first_token_ms":249,"total_runtime_ms":1355,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0253,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.109,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1289,"p95_latency_ms":1575,"p99_latency_ms":1689,"time_to_first_token_ms":257,"total_runtime_ms":1432,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0209,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9821,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":816,"p95_latency_ms":997,"p99_latency_ms":1070,"time_to_first_token_ms":192,"total_runtime_ms":906,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0249,"confidence_interval":0.9352,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9632,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":790,"p95_latency_ms":965,"p99_latency_ms":1036,"time_to_first_token_ms":193,"total_runtime_ms":878,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0278,"confidence_interval":0.9355,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8531,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":796,"p95_latency_ms":972,"p99_latency_ms":1043,"time_to_first_token_ms":203,"total_runtime_ms":884,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0286,"confidence_interval":0.9401,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7416,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":762,"p95_latency_ms":931,"p99_latency_ms":999,"time_to_first_token_ms":183,"total_runtime_ms":847,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0274,"confidence_interval":0.9432,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6849,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":210,"total_runtime_ms":849,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0225,"confidence_interval":0.9424,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8369,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1053,"time_to_first_token_ms":178,"total_runtime_ms":892,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9493,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7231,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":180,"total_runtime_ms":889,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0255,"confidence_interval":0.9487,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2326,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":808,"p95_latency_ms":987,"p99_latency_ms":1059,"time_to_first_token_ms":185,"total_runtime_ms":898,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0165,"confidence_interval":0.9357,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9224,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":758,"p95_latency_ms":926,"p99_latency_ms":993,"time_to_first_token_ms":200,"total_runtime_ms":842,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0189,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9349,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:7b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1009,"time_to_first_token_ms":202,"total_runtime_ms":855,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.023,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3281,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1763,"p95_latency_ms":2155,"p99_latency_ms":2312,"time_to_first_token_ms":301,"total_runtime_ms":1959,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0279,"confidence_interval":0.9375,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4057,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1781,"p95_latency_ms":2177,"p99_latency_ms":2335,"time_to_first_token_ms":316,"total_runtime_ms":1979,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0246,"confidence_interval":0.9437,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.87,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1851,"p95_latency_ms":2263,"p99_latency_ms":2427,"time_to_first_token_ms":310,"total_runtime_ms":2057,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0222,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4499,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1838,"p95_latency_ms":2247,"p99_latency_ms":2410,"time_to_first_token_ms":335,"total_runtime_ms":2043,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.011,"confidence_interval":0.9384,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.209,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1766,"p95_latency_ms":2159,"p99_latency_ms":2316,"time_to_first_token_ms":323,"total_runtime_ms":1962,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0143,"confidence_interval":0.9315,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8058,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1760,"p95_latency_ms":2152,"p99_latency_ms":2308,"time_to_first_token_ms":318,"total_runtime_ms":1956,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0114,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7441,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1822,"p95_latency_ms":2226,"p99_latency_ms":2388,"time_to_first_token_ms":302,"total_runtime_ms":2024,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.024,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2558,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1799,"p95_latency_ms":2199,"p99_latency_ms":2358,"time_to_first_token_ms":302,"total_runtime_ms":1999,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0233,"confidence_interval":0.9449,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1221,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1863,"p95_latency_ms":2277,"p99_latency_ms":2442,"time_to_first_token_ms":300,"total_runtime_ms":2070,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0178,"confidence_interval":0.9468,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7367,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1842,"p95_latency_ms":2251,"p99_latency_ms":2415,"time_to_first_token_ms":320,"total_runtime_ms":2047,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0185,"confidence_interval":0.9344,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5373,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1134,"p95_latency_ms":1386,"p99_latency_ms":1487,"time_to_first_token_ms":271,"total_runtime_ms":1260,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0238,"confidence_interval":0.9402,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.531,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1090,"p95_latency_ms":1333,"p99_latency_ms":1430,"time_to_first_token_ms":277,"total_runtime_ms":1212,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7511,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1118,"p95_latency_ms":1367,"p99_latency_ms":1466,"time_to_first_token_ms":270,"total_runtime_ms":1242,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0146,"confidence_interval":0.9445,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8975,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1159,"p95_latency_ms":1416,"p99_latency_ms":1519,"time_to_first_token_ms":283,"total_runtime_ms":1288,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0165,"confidence_interval":0.9387,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7291,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1111,"p95_latency_ms":1358,"p99_latency_ms":1456,"time_to_first_token_ms":271,"total_runtime_ms":1234,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0295,"confidence_interval":0.939,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1133,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1098,"p95_latency_ms":1342,"p99_latency_ms":1439,"time_to_first_token_ms":266,"total_runtime_ms":1220,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0188,"confidence_interval":0.9347,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2818,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1062,"p95_latency_ms":1298,"p99_latency_ms":1392,"time_to_first_token_ms":274,"total_runtime_ms":1180,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0258,"confidence_interval":0.9349,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4496,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1074,"p95_latency_ms":1313,"p99_latency_ms":1408,"time_to_first_token_ms":264,"total_runtime_ms":1194,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0107,"confidence_interval":0.9381,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.626,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1130,"p95_latency_ms":1381,"p99_latency_ms":1481,"time_to_first_token_ms":280,"total_runtime_ms":1255,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9395,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9113,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"analyze_reasoning|high|qwen2:14b|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1146,"p95_latency_ms":1401,"p99_latency_ms":1503,"time_to_first_token_ms":267,"total_runtime_ms":1274,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0105,"confidence_interval":0.9336,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5336,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1024,"p95_latency_ms":1252,"p99_latency_ms":1343,"time_to_first_token_ms":217,"total_runtime_ms":1138,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0217,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5747,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":958,"p95_latency_ms":1170,"p99_latency_ms":1255,"time_to_first_token_ms":208,"total_runtime_ms":1064,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0275,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2488,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":989,"p95_latency_ms":1208,"p99_latency_ms":1296,"time_to_first_token_ms":199,"total_runtime_ms":1099,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0157,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8959,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1011,"p95_latency_ms":1236,"p99_latency_ms":1325,"time_to_first_token_ms":216,"total_runtime_ms":1123,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0171,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6134,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":982,"p95_latency_ms":1200,"p99_latency_ms":1287,"time_to_first_token_ms":207,"total_runtime_ms":1091,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0179,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9265,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1005,"p95_latency_ms":1229,"p99_latency_ms":1318,"time_to_first_token_ms":201,"total_runtime_ms":1117,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0146,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7035,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":981,"p95_latency_ms":1199,"p99_latency_ms":1287,"time_to_first_token_ms":205,"total_runtime_ms":1090,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.021,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1449,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":993,"p95_latency_ms":1214,"p99_latency_ms":1302,"time_to_first_token_ms":220,"total_runtime_ms":1104,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0163,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5828,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1000,"p95_latency_ms":1222,"p99_latency_ms":1311,"time_to_first_token_ms":218,"total_runtime_ms":1111,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0206,"confidence_interval":0.9357,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3558,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":992,"p95_latency_ms":1212,"p99_latency_ms":1300,"time_to_first_token_ms":201,"total_runtime_ms":1102,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.017,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4858,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":667,"p95_latency_ms":816,"p99_latency_ms":875,"time_to_first_token_ms":153,"total_runtime_ms":741,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0102,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0052,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":636,"p95_latency_ms":777,"p99_latency_ms":834,"time_to_first_token_ms":149,"total_runtime_ms":707,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0166,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3917,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":651,"p95_latency_ms":796,"p99_latency_ms":854,"time_to_first_token_ms":154,"total_runtime_ms":724,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0224,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4731,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":609,"p95_latency_ms":744,"p99_latency_ms":798,"time_to_first_token_ms":151,"total_runtime_ms":676,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0156,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.055,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":575,"p95_latency_ms":702,"p99_latency_ms":753,"time_to_first_token_ms":174,"total_runtime_ms":638,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0206,"confidence_interval":0.9448,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7568,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":603,"p95_latency_ms":737,"p99_latency_ms":791,"time_to_first_token_ms":164,"total_runtime_ms":670,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0154,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.891,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":636,"p95_latency_ms":777,"p99_latency_ms":834,"time_to_first_token_ms":164,"total_runtime_ms":707,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0156,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9656,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":636,"p95_latency_ms":778,"p99_latency_ms":834,"time_to_first_token_ms":170,"total_runtime_ms":707,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0253,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9369,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":564,"p95_latency_ms":690,"p99_latency_ms":740,"time_to_first_token_ms":177,"total_runtime_ms":627,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0174,"confidence_interval":0.9454,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4669,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":843,"time_to_first_token_ms":161,"total_runtime_ms":714,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.011,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.597,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1356,"p95_latency_ms":1657,"p99_latency_ms":1777,"time_to_first_token_ms":263,"total_runtime_ms":1506,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0259,"confidence_interval":0.9348,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0301,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1329,"p95_latency_ms":1624,"p99_latency_ms":1743,"time_to_first_token_ms":264,"total_runtime_ms":1477,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9412,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6314,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1320,"p95_latency_ms":1613,"p99_latency_ms":1731,"time_to_first_token_ms":261,"total_runtime_ms":1467,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0234,"confidence_interval":0.943,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4591,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1297,"p95_latency_ms":1585,"p99_latency_ms":1701,"time_to_first_token_ms":267,"total_runtime_ms":1441,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0139,"confidence_interval":0.9334,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4718,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1294,"p95_latency_ms":1582,"p99_latency_ms":1697,"time_to_first_token_ms":260,"total_runtime_ms":1438,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0272,"confidence_interval":0.9379,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9937,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1347,"p95_latency_ms":1646,"p99_latency_ms":1766,"time_to_first_token_ms":259,"total_runtime_ms":1497,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0199,"confidence_interval":0.9422,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8808,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1341,"p95_latency_ms":1639,"p99_latency_ms":1758,"time_to_first_token_ms":249,"total_runtime_ms":1490,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0209,"confidence_interval":0.9401,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6819,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1352,"p95_latency_ms":1652,"p99_latency_ms":1773,"time_to_first_token_ms":254,"total_runtime_ms":1502,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.9384,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6924,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1329,"p95_latency_ms":1625,"p99_latency_ms":1743,"time_to_first_token_ms":248,"total_runtime_ms":1477,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0118,"confidence_interval":0.9361,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8453,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1336,"p95_latency_ms":1633,"p99_latency_ms":1752,"time_to_first_token_ms":266,"total_runtime_ms":1485,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0251,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9093,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":813,"p95_latency_ms":993,"p99_latency_ms":1066,"time_to_first_token_ms":215,"total_runtime_ms":903,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0187,"confidence_interval":0.9405,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7579,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":832,"p95_latency_ms":1017,"p99_latency_ms":1091,"time_to_first_token_ms":221,"total_runtime_ms":925,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.9366,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4355,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":836,"p95_latency_ms":1022,"p99_latency_ms":1096,"time_to_first_token_ms":220,"total_runtime_ms":929,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0143,"confidence_interval":0.9365,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8549,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":823,"p95_latency_ms":1006,"p99_latency_ms":1079,"time_to_first_token_ms":191,"total_runtime_ms":914,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0107,"confidence_interval":0.9403,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3995,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":842,"p95_latency_ms":1029,"p99_latency_ms":1104,"time_to_first_token_ms":202,"total_runtime_ms":936,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0145,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.078,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":847,"p95_latency_ms":1036,"p99_latency_ms":1111,"time_to_first_token_ms":206,"total_runtime_ms":942,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0231,"confidence_interval":0.9478,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5913,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":802,"p95_latency_ms":980,"p99_latency_ms":1052,"time_to_first_token_ms":223,"total_runtime_ms":891,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0198,"confidence_interval":0.9365,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5145,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":843,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":216,"total_runtime_ms":937,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0283,"confidence_interval":0.9403,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4953,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":842,"p95_latency_ms":1029,"p99_latency_ms":1104,"time_to_first_token_ms":204,"total_runtime_ms":936,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0169,"confidence_interval":0.934,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6343,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":835,"p95_latency_ms":1021,"p99_latency_ms":1095,"time_to_first_token_ms":195,"total_runtime_ms":928,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0129,"confidence_interval":0.9344,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5609,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2577,"p95_latency_ms":3150,"p99_latency_ms":3379,"time_to_first_token_ms":429,"total_runtime_ms":2864,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0223,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6978,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2625,"p95_latency_ms":3209,"p99_latency_ms":3442,"time_to_first_token_ms":415,"total_runtime_ms":2917,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.027,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6794,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2629,"p95_latency_ms":3213,"p99_latency_ms":3447,"time_to_first_token_ms":422,"total_runtime_ms":2921,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0237,"confidence_interval":0.9325,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8154,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2624,"p95_latency_ms":3208,"p99_latency_ms":3441,"time_to_first_token_ms":421,"total_runtime_ms":2916,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9346,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5143,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2533,"p95_latency_ms":3096,"p99_latency_ms":3321,"time_to_first_token_ms":413,"total_runtime_ms":2815,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0172,"confidence_interval":0.9365,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9105,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2624,"p95_latency_ms":3207,"p99_latency_ms":3441,"time_to_first_token_ms":407,"total_runtime_ms":2916,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0134,"confidence_interval":0.9479,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0724,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2612,"p95_latency_ms":3193,"p99_latency_ms":3425,"time_to_first_token_ms":431,"total_runtime_ms":2902,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0118,"confidence_interval":0.9423,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3519,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2567,"p95_latency_ms":3137,"p99_latency_ms":3365,"time_to_first_token_ms":411,"total_runtime_ms":2852,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0122,"confidence_interval":0.9419,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5726,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2588,"p95_latency_ms":3163,"p99_latency_ms":3393,"time_to_first_token_ms":419,"total_runtime_ms":2875,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0117,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7923,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2572,"p95_latency_ms":3143,"p99_latency_ms":3372,"time_to_first_token_ms":409,"total_runtime_ms":2857,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0203,"confidence_interval":0.9309,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3325,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1574,"p95_latency_ms":1924,"p99_latency_ms":2064,"time_to_first_token_ms":368,"total_runtime_ms":1749,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0144,"confidence_interval":0.9368,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3776,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1555,"p95_latency_ms":1901,"p99_latency_ms":2039,"time_to_first_token_ms":373,"total_runtime_ms":1728,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0136,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5334,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1519,"p95_latency_ms":1856,"p99_latency_ms":1991,"time_to_first_token_ms":379,"total_runtime_ms":1688,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0105,"confidence_interval":0.9398,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9439,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1533,"p95_latency_ms":1874,"p99_latency_ms":2010,"time_to_first_token_ms":375,"total_runtime_ms":1704,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0174,"confidence_interval":0.9393,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1901,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1572,"p95_latency_ms":1922,"p99_latency_ms":2061,"time_to_first_token_ms":391,"total_runtime_ms":1747,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0175,"confidence_interval":0.9398,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0192,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1550,"p95_latency_ms":1895,"p99_latency_ms":2032,"time_to_first_token_ms":361,"total_runtime_ms":1722,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0197,"confidence_interval":0.9358,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7168,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1607,"p95_latency_ms":1964,"p99_latency_ms":2106,"time_to_first_token_ms":377,"total_runtime_ms":1785,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0248,"confidence_interval":0.9386,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4797,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1573,"p95_latency_ms":1922,"p99_latency_ms":2062,"time_to_first_token_ms":391,"total_runtime_ms":1748,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0182,"confidence_interval":0.9457,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7973,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1549,"p95_latency_ms":1893,"p99_latency_ms":2031,"time_to_first_token_ms":392,"total_runtime_ms":1721,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0113,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0316,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:7b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1571,"p95_latency_ms":1920,"p99_latency_ms":2060,"time_to_first_token_ms":391,"total_runtime_ms":1746,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0136,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5285,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3804,"p95_latency_ms":4649,"p99_latency_ms":4987,"time_to_first_token_ms":576,"total_runtime_ms":4227,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0104,"confidence_interval":0.9322,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.274,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3804,"p95_latency_ms":4650,"p99_latency_ms":4988,"time_to_first_token_ms":576,"total_runtime_ms":4227,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0261,"confidence_interval":0.936,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0214,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3797,"p95_latency_ms":4640,"p99_latency_ms":4978,"time_to_first_token_ms":596,"total_runtime_ms":4219,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0127,"confidence_interval":0.9358,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1915,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3772,"p95_latency_ms":4610,"p99_latency_ms":4946,"time_to_first_token_ms":579,"total_runtime_ms":4191,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0177,"confidence_interval":0.9327,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2193,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3844,"p95_latency_ms":4699,"p99_latency_ms":5040,"time_to_first_token_ms":581,"total_runtime_ms":4272,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0224,"confidence_interval":0.9423,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.482,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3745,"p95_latency_ms":4577,"p99_latency_ms":4910,"time_to_first_token_ms":597,"total_runtime_ms":4161,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0201,"confidence_interval":0.9467,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3024,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3834,"p95_latency_ms":4686,"p99_latency_ms":5027,"time_to_first_token_ms":582,"total_runtime_ms":4260,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.017,"confidence_interval":0.9315,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7096,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3840,"p95_latency_ms":4693,"p99_latency_ms":5034,"time_to_first_token_ms":585,"total_runtime_ms":4266,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0159,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4499,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3757,"p95_latency_ms":4592,"p99_latency_ms":4926,"time_to_first_token_ms":604,"total_runtime_ms":4174,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0105,"confidence_interval":0.9381,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0884,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3745,"p95_latency_ms":4577,"p99_latency_ms":4910,"time_to_first_token_ms":582,"total_runtime_ms":4161,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0287,"confidence_interval":0.9449,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3067,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2342,"p95_latency_ms":2862,"p99_latency_ms":3070,"time_to_first_token_ms":556,"total_runtime_ms":2602,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0142,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2896,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2345,"p95_latency_ms":2867,"p99_latency_ms":3075,"time_to_first_token_ms":550,"total_runtime_ms":2606,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0147,"confidence_interval":0.945,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.991,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2247,"p95_latency_ms":2746,"p99_latency_ms":2946,"time_to_first_token_ms":561,"total_runtime_ms":2497,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0262,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0204,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2313,"p95_latency_ms":2827,"p99_latency_ms":3032,"time_to_first_token_ms":556,"total_runtime_ms":2570,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.021,"confidence_interval":0.9461,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7352,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2322,"p95_latency_ms":2838,"p99_latency_ms":3044,"time_to_first_token_ms":531,"total_runtime_ms":2580,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0132,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1256,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2286,"p95_latency_ms":2793,"p99_latency_ms":2997,"time_to_first_token_ms":535,"total_runtime_ms":2540,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0257,"confidence_interval":0.9313,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0517,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2310,"p95_latency_ms":2823,"p99_latency_ms":3029,"time_to_first_token_ms":546,"total_runtime_ms":2567,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0207,"confidence_interval":0.9433,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3276,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2330,"p95_latency_ms":2848,"p99_latency_ms":3056,"time_to_first_token_ms":536,"total_runtime_ms":2589,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0275,"confidence_interval":0.9486,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7204,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2308,"p95_latency_ms":2820,"p99_latency_ms":3025,"time_to_first_token_ms":556,"total_runtime_ms":2564,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0236,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7629,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|low|qwen2:14b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2301,"p95_latency_ms":2812,"p99_latency_ms":3017,"time_to_first_token_ms":529,"total_runtime_ms":2557,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.9312,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7124,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":803,"p95_latency_ms":982,"p99_latency_ms":1053,"time_to_first_token_ms":183,"total_runtime_ms":892,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0105,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8727,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":171,"total_runtime_ms":889,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0118,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8601,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":719,"p95_latency_ms":879,"p99_latency_ms":943,"time_to_first_token_ms":174,"total_runtime_ms":799,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0122,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3584,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":806,"p95_latency_ms":985,"p99_latency_ms":1057,"time_to_first_token_ms":179,"total_runtime_ms":896,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.016,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6019,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":801,"p95_latency_ms":979,"p99_latency_ms":1050,"time_to_first_token_ms":164,"total_runtime_ms":890,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0242,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1902,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":183,"total_runtime_ms":895,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0148,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":753,"p95_latency_ms":920,"p99_latency_ms":987,"time_to_first_token_ms":167,"total_runtime_ms":836,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0145,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7688,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":756,"p95_latency_ms":924,"p99_latency_ms":991,"time_to_first_token_ms":187,"total_runtime_ms":840,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0173,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.046,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":743,"p95_latency_ms":908,"p99_latency_ms":974,"time_to_first_token_ms":160,"total_runtime_ms":826,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0115,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5252,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1002,"time_to_first_token_ms":169,"total_runtime_ms":850,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0103,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.707,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":519,"p95_latency_ms":635,"p99_latency_ms":681,"time_to_first_token_ms":143,"total_runtime_ms":577,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0234,"confidence_interval":0.93,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8611,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":496,"p95_latency_ms":606,"p99_latency_ms":650,"time_to_first_token_ms":133,"total_runtime_ms":551,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0127,"confidence_interval":0.9381,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9153,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":669,"time_to_first_token_ms":142,"total_runtime_ms":567,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0129,"confidence_interval":0.9381,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1943,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":577,"time_to_first_token_ms":145,"total_runtime_ms":489,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.02,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.27,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":430,"p95_latency_ms":526,"p99_latency_ms":564,"time_to_first_token_ms":113,"total_runtime_ms":478,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0173,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3153,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":490,"p95_latency_ms":599,"p99_latency_ms":643,"time_to_first_token_ms":145,"total_runtime_ms":545,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0256,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3424,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":460,"p95_latency_ms":563,"p99_latency_ms":604,"time_to_first_token_ms":141,"total_runtime_ms":511,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0215,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3735,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":495,"p95_latency_ms":605,"p99_latency_ms":649,"time_to_first_token_ms":125,"total_runtime_ms":550,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0182,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0954,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":120,"total_runtime_ms":510,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0139,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4858,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":465,"p95_latency_ms":568,"p99_latency_ms":610,"time_to_first_token_ms":143,"total_runtime_ms":517,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0126,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1334,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1002,"p95_latency_ms":1225,"p99_latency_ms":1314,"time_to_first_token_ms":196,"total_runtime_ms":1113,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9482,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9645,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1045,"p95_latency_ms":1277,"p99_latency_ms":1370,"time_to_first_token_ms":213,"total_runtime_ms":1161,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0217,"confidence_interval":0.9369,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2671,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1011,"p95_latency_ms":1236,"p99_latency_ms":1326,"time_to_first_token_ms":191,"total_runtime_ms":1124,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0135,"confidence_interval":0.9418,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.374,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":949,"p95_latency_ms":1160,"p99_latency_ms":1244,"time_to_first_token_ms":206,"total_runtime_ms":1055,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0208,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5108,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1007,"p95_latency_ms":1231,"p99_latency_ms":1320,"time_to_first_token_ms":208,"total_runtime_ms":1119,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9806,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1021,"p95_latency_ms":1248,"p99_latency_ms":1339,"time_to_first_token_ms":195,"total_runtime_ms":1134,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0155,"confidence_interval":0.9313,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4179,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":999,"p95_latency_ms":1221,"p99_latency_ms":1310,"time_to_first_token_ms":201,"total_runtime_ms":1110,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0222,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.384,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1015,"p95_latency_ms":1241,"p99_latency_ms":1331,"time_to_first_token_ms":206,"total_runtime_ms":1128,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0269,"confidence_interval":0.9318,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.146,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":964,"p95_latency_ms":1179,"p99_latency_ms":1264,"time_to_first_token_ms":205,"total_runtime_ms":1072,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0128,"confidence_interval":0.9394,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5194,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":975,"p95_latency_ms":1192,"p99_latency_ms":1278,"time_to_first_token_ms":222,"total_runtime_ms":1083,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0193,"confidence_interval":0.9436,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.466,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":640,"p95_latency_ms":783,"p99_latency_ms":840,"time_to_first_token_ms":145,"total_runtime_ms":712,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0124,"confidence_interval":0.9472,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4103,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":673,"p95_latency_ms":823,"p99_latency_ms":883,"time_to_first_token_ms":176,"total_runtime_ms":748,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0183,"confidence_interval":0.9371,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2187,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":664,"p95_latency_ms":812,"p99_latency_ms":871,"time_to_first_token_ms":158,"total_runtime_ms":738,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0118,"confidence_interval":0.9444,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.876,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":601,"p95_latency_ms":735,"p99_latency_ms":788,"time_to_first_token_ms":178,"total_runtime_ms":668,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.9345,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.671,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":638,"p95_latency_ms":779,"p99_latency_ms":836,"time_to_first_token_ms":152,"total_runtime_ms":709,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0103,"confidence_interval":0.9477,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6894,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":671,"p95_latency_ms":820,"p99_latency_ms":879,"time_to_first_token_ms":150,"total_runtime_ms":745,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.013,"confidence_interval":0.9323,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.117,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":173,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0295,"confidence_interval":0.931,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3737,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":655,"p95_latency_ms":800,"p99_latency_ms":858,"time_to_first_token_ms":166,"total_runtime_ms":727,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.9355,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3789,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":622,"p95_latency_ms":760,"p99_latency_ms":815,"time_to_first_token_ms":180,"total_runtime_ms":691,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0266,"confidence_interval":0.9489,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9622,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":851,"time_to_first_token_ms":173,"total_runtime_ms":721,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0164,"confidence_interval":0.9432,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0794,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1966,"p95_latency_ms":2403,"p99_latency_ms":2578,"time_to_first_token_ms":336,"total_runtime_ms":2185,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0173,"confidence_interval":0.9434,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4887,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1949,"p95_latency_ms":2382,"p99_latency_ms":2555,"time_to_first_token_ms":323,"total_runtime_ms":2165,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0183,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1805,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1956,"p95_latency_ms":2390,"p99_latency_ms":2564,"time_to_first_token_ms":320,"total_runtime_ms":2173,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0185,"confidence_interval":0.9467,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7999,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1972,"p95_latency_ms":2411,"p99_latency_ms":2586,"time_to_first_token_ms":328,"total_runtime_ms":2192,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0201,"confidence_interval":0.9453,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8925,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1908,"p95_latency_ms":2332,"p99_latency_ms":2502,"time_to_first_token_ms":345,"total_runtime_ms":2120,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0238,"confidence_interval":0.9357,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.997,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1974,"p95_latency_ms":2413,"p99_latency_ms":2589,"time_to_first_token_ms":343,"total_runtime_ms":2194,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0266,"confidence_interval":0.9371,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9186,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1904,"p95_latency_ms":2327,"p99_latency_ms":2496,"time_to_first_token_ms":320,"total_runtime_ms":2115,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0119,"confidence_interval":0.9313,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5778,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1919,"p95_latency_ms":2346,"p99_latency_ms":2516,"time_to_first_token_ms":331,"total_runtime_ms":2132,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0182,"confidence_interval":0.9325,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1895,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1993,"p95_latency_ms":2436,"p99_latency_ms":2613,"time_to_first_token_ms":346,"total_runtime_ms":2215,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0124,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1015,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1935,"p95_latency_ms":2365,"p99_latency_ms":2537,"time_to_first_token_ms":340,"total_runtime_ms":2150,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0298,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7005,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1148,"p95_latency_ms":1403,"p99_latency_ms":1505,"time_to_first_token_ms":287,"total_runtime_ms":1275,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0189,"confidence_interval":0.9369,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9982,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1226,"p95_latency_ms":1498,"p99_latency_ms":1607,"time_to_first_token_ms":301,"total_runtime_ms":1362,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0102,"confidence_interval":0.9337,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7695,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1173,"p95_latency_ms":1434,"p99_latency_ms":1539,"time_to_first_token_ms":292,"total_runtime_ms":1304,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.027,"confidence_interval":0.9309,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.673,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1140,"p95_latency_ms":1394,"p99_latency_ms":1495,"time_to_first_token_ms":291,"total_runtime_ms":1267,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.012,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3819,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1242,"p95_latency_ms":1518,"p99_latency_ms":1629,"time_to_first_token_ms":296,"total_runtime_ms":1380,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0115,"confidence_interval":0.9485,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.533,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1139,"p95_latency_ms":1392,"p99_latency_ms":1493,"time_to_first_token_ms":295,"total_runtime_ms":1266,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9426,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2357,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1167,"p95_latency_ms":1427,"p99_latency_ms":1530,"time_to_first_token_ms":288,"total_runtime_ms":1297,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0128,"confidence_interval":0.9414,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1251,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1189,"p95_latency_ms":1453,"p99_latency_ms":1559,"time_to_first_token_ms":293,"total_runtime_ms":1321,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.023,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6661,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1205,"p95_latency_ms":1473,"p99_latency_ms":1580,"time_to_first_token_ms":300,"total_runtime_ms":1339,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0129,"confidence_interval":0.9488,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8475,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1196,"p95_latency_ms":1462,"p99_latency_ms":1568,"time_to_first_token_ms":299,"total_runtime_ms":1329,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0202,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.988,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2832,"p95_latency_ms":3461,"p99_latency_ms":3713,"time_to_first_token_ms":447,"total_runtime_ms":3146,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0131,"confidence_interval":0.9423,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5021,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2860,"p95_latency_ms":3495,"p99_latency_ms":3750,"time_to_first_token_ms":456,"total_runtime_ms":3178,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.9467,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9537,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2825,"p95_latency_ms":3453,"p99_latency_ms":3705,"time_to_first_token_ms":449,"total_runtime_ms":3139,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.022,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6731,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2858,"p95_latency_ms":3493,"p99_latency_ms":3747,"time_to_first_token_ms":450,"total_runtime_ms":3176,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0229,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.389,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2818,"p95_latency_ms":3445,"p99_latency_ms":3695,"time_to_first_token_ms":446,"total_runtime_ms":3131,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0249,"confidence_interval":0.937,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9575,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2882,"p95_latency_ms":3523,"p99_latency_ms":3779,"time_to_first_token_ms":462,"total_runtime_ms":3203,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0182,"confidence_interval":0.947,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9525,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2903,"p95_latency_ms":3548,"p99_latency_ms":3806,"time_to_first_token_ms":450,"total_runtime_ms":3226,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0297,"confidence_interval":0.9412,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0353,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2858,"p95_latency_ms":3493,"p99_latency_ms":3747,"time_to_first_token_ms":456,"total_runtime_ms":3175,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0214,"confidence_interval":0.9302,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3115,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2885,"p95_latency_ms":3526,"p99_latency_ms":3782,"time_to_first_token_ms":452,"total_runtime_ms":3205,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.025,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7353,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2872,"p95_latency_ms":3510,"p99_latency_ms":3765,"time_to_first_token_ms":446,"total_runtime_ms":3191,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0149,"confidence_interval":0.9361,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0145,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1766,"p95_latency_ms":2158,"p99_latency_ms":2315,"time_to_first_token_ms":401,"total_runtime_ms":1962,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0248,"confidence_interval":0.9385,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2123,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1737,"p95_latency_ms":2122,"p99_latency_ms":2277,"time_to_first_token_ms":430,"total_runtime_ms":1929,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0161,"confidence_interval":0.944,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8346,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1704,"p95_latency_ms":2083,"p99_latency_ms":2234,"time_to_first_token_ms":401,"total_runtime_ms":1893,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0273,"confidence_interval":0.9313,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4118,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1698,"p95_latency_ms":2075,"p99_latency_ms":2226,"time_to_first_token_ms":409,"total_runtime_ms":1887,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.011,"confidence_interval":0.9352,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7819,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1747,"p95_latency_ms":2135,"p99_latency_ms":2291,"time_to_first_token_ms":419,"total_runtime_ms":1941,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0262,"confidence_interval":0.944,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3765,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1757,"p95_latency_ms":2148,"p99_latency_ms":2304,"time_to_first_token_ms":428,"total_runtime_ms":1952,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0164,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0882,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1788,"p95_latency_ms":2186,"p99_latency_ms":2345,"time_to_first_token_ms":413,"total_runtime_ms":1987,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0265,"confidence_interval":0.9465,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3401,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1690,"p95_latency_ms":2066,"p99_latency_ms":2216,"time_to_first_token_ms":410,"total_runtime_ms":1878,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9448,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8735,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1729,"p95_latency_ms":2114,"p99_latency_ms":2267,"time_to_first_token_ms":402,"total_runtime_ms":1922,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9439,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7355,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1701,"p95_latency_ms":2079,"p99_latency_ms":2230,"time_to_first_token_ms":422,"total_runtime_ms":1890,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.021,"confidence_interval":0.9354,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6625,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":609,"p95_latency_ms":745,"p99_latency_ms":799,"time_to_first_token_ms":153,"total_runtime_ms":677,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0235,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9313,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":596,"p95_latency_ms":728,"p99_latency_ms":781,"time_to_first_token_ms":149,"total_runtime_ms":662,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3165,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":571,"p95_latency_ms":698,"p99_latency_ms":749,"time_to_first_token_ms":158,"total_runtime_ms":634,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0168,"confidence_interval":0.9331,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2634,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":566,"p95_latency_ms":691,"p99_latency_ms":742,"time_to_first_token_ms":147,"total_runtime_ms":629,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0199,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7539,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":583,"p95_latency_ms":713,"p99_latency_ms":764,"time_to_first_token_ms":147,"total_runtime_ms":648,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.023,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4823,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":608,"p95_latency_ms":744,"p99_latency_ms":798,"time_to_first_token_ms":136,"total_runtime_ms":676,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0224,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0028,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":605,"p95_latency_ms":739,"p99_latency_ms":793,"time_to_first_token_ms":140,"total_runtime_ms":672,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0124,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4269,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":605,"p95_latency_ms":739,"p99_latency_ms":793,"time_to_first_token_ms":157,"total_runtime_ms":672,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0105,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5096,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":594,"p95_latency_ms":726,"p99_latency_ms":778,"time_to_first_token_ms":163,"total_runtime_ms":660,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0175,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.007,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":575,"p95_latency_ms":703,"p99_latency_ms":754,"time_to_first_token_ms":164,"total_runtime_ms":639,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1849,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":393,"p95_latency_ms":481,"p99_latency_ms":515,"time_to_first_token_ms":91,"total_runtime_ms":437,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0203,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1773,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":343,"p95_latency_ms":420,"p99_latency_ms":450,"time_to_first_token_ms":88,"total_runtime_ms":382,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0193,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1584,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":548,"time_to_first_token_ms":119,"total_runtime_ms":464,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0103,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5188,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":352,"p95_latency_ms":430,"p99_latency_ms":461,"time_to_first_token_ms":107,"total_runtime_ms":391,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0128,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3862,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":317,"p95_latency_ms":388,"p99_latency_ms":416,"time_to_first_token_ms":91,"total_runtime_ms":353,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0226,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7331,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":374,"p95_latency_ms":457,"p99_latency_ms":491,"time_to_first_token_ms":90,"total_runtime_ms":416,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0193,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6378,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":361,"p95_latency_ms":441,"p99_latency_ms":473,"time_to_first_token_ms":111,"total_runtime_ms":401,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0424,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":401,"p95_latency_ms":491,"p99_latency_ms":526,"time_to_first_token_ms":88,"total_runtime_ms":446,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0199,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8439,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":329,"p95_latency_ms":402,"p99_latency_ms":431,"time_to_first_token_ms":117,"total_runtime_ms":366,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0245,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4274,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":379,"p95_latency_ms":463,"p99_latency_ms":497,"time_to_first_token_ms":91,"total_runtime_ms":421,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0253,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0019,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":733,"p95_latency_ms":896,"p99_latency_ms":961,"time_to_first_token_ms":168,"total_runtime_ms":815,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0105,"confidence_interval":0.9408,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5949,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":754,"p95_latency_ms":922,"p99_latency_ms":989,"time_to_first_token_ms":191,"total_runtime_ms":838,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0271,"confidence_interval":0.9455,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2041,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":748,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":186,"total_runtime_ms":831,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0216,"confidence_interval":0.9412,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5141,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":777,"p95_latency_ms":950,"p99_latency_ms":1019,"time_to_first_token_ms":162,"total_runtime_ms":863,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0154,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5828,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":743,"p95_latency_ms":909,"p99_latency_ms":975,"time_to_first_token_ms":192,"total_runtime_ms":826,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0226,"confidence_interval":0.9479,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.661,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":743,"p95_latency_ms":909,"p99_latency_ms":975,"time_to_first_token_ms":189,"total_runtime_ms":826,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0251,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6656,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":189,"total_runtime_ms":895,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0265,"confidence_interval":0.9398,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0532,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1010,"time_to_first_token_ms":160,"total_runtime_ms":856,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0248,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8387,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":792,"p95_latency_ms":968,"p99_latency_ms":1039,"time_to_first_token_ms":177,"total_runtime_ms":880,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0101,"confidence_interval":0.9398,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3631,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":948,"time_to_first_token_ms":180,"total_runtime_ms":804,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":41,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0171,"confidence_interval":0.9413,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5277,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":492,"p95_latency_ms":601,"p99_latency_ms":645,"time_to_first_token_ms":114,"total_runtime_ms":547,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0277,"confidence_interval":0.9372,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7488,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":501,"p95_latency_ms":612,"p99_latency_ms":656,"time_to_first_token_ms":125,"total_runtime_ms":556,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0258,"confidence_interval":0.9434,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1746,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":508,"p95_latency_ms":621,"p99_latency_ms":666,"time_to_first_token_ms":142,"total_runtime_ms":565,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0111,"confidence_interval":0.9472,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.558,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":668,"time_to_first_token_ms":146,"total_runtime_ms":566,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0133,"confidence_interval":0.9429,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1101,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":496,"p95_latency_ms":606,"p99_latency_ms":651,"time_to_first_token_ms":144,"total_runtime_ms":551,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0263,"confidence_interval":0.9406,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9333,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":478,"p95_latency_ms":584,"p99_latency_ms":626,"time_to_first_token_ms":139,"total_runtime_ms":531,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0159,"confidence_interval":0.9402,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4901,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":134,"total_runtime_ms":586,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0255,"confidence_interval":0.9429,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4834,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":444,"p95_latency_ms":542,"p99_latency_ms":582,"time_to_first_token_ms":121,"total_runtime_ms":493,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0189,"confidence_interval":0.9465,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2246,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":580,"p99_latency_ms":622,"time_to_first_token_ms":131,"total_runtime_ms":527,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0157,"confidence_interval":0.9352,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0129,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":121,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":41,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0162,"confidence_interval":0.9339,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.984,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1476,"p95_latency_ms":1804,"p99_latency_ms":1935,"time_to_first_token_ms":289,"total_runtime_ms":1640,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0233,"confidence_interval":0.944,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2609,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1451,"p95_latency_ms":1773,"p99_latency_ms":1902,"time_to_first_token_ms":277,"total_runtime_ms":1612,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.03,"confidence_interval":0.9453,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3525,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1463,"p95_latency_ms":1788,"p99_latency_ms":1918,"time_to_first_token_ms":271,"total_runtime_ms":1626,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0174,"confidence_interval":0.9357,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1898,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1504,"p95_latency_ms":1838,"p99_latency_ms":1972,"time_to_first_token_ms":282,"total_runtime_ms":1671,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0263,"confidence_interval":0.937,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7123,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1422,"p95_latency_ms":1738,"p99_latency_ms":1864,"time_to_first_token_ms":277,"total_runtime_ms":1580,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0111,"confidence_interval":0.9339,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7745,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1496,"p95_latency_ms":1828,"p99_latency_ms":1961,"time_to_first_token_ms":267,"total_runtime_ms":1662,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0119,"confidence_interval":0.9487,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3524,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1506,"p95_latency_ms":1841,"p99_latency_ms":1975,"time_to_first_token_ms":277,"total_runtime_ms":1673,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0106,"confidence_interval":0.9459,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5118,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1458,"p95_latency_ms":1782,"p99_latency_ms":1912,"time_to_first_token_ms":262,"total_runtime_ms":1620,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0128,"confidence_interval":0.9482,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.187,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1459,"p95_latency_ms":1783,"p99_latency_ms":1913,"time_to_first_token_ms":258,"total_runtime_ms":1621,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0264,"confidence_interval":0.9383,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0211,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1522,"p95_latency_ms":1861,"p99_latency_ms":1996,"time_to_first_token_ms":271,"total_runtime_ms":1691,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":81,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0119,"confidence_interval":0.9326,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4222,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":888,"p95_latency_ms":1086,"p99_latency_ms":1165,"time_to_first_token_ms":210,"total_runtime_ms":987,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0284,"confidence_interval":0.9356,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8169,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":953,"p95_latency_ms":1164,"p99_latency_ms":1249,"time_to_first_token_ms":220,"total_runtime_ms":1058,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.019,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4402,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":877,"p95_latency_ms":1072,"p99_latency_ms":1149,"time_to_first_token_ms":220,"total_runtime_ms":974,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0176,"confidence_interval":0.942,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5644,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":883,"p95_latency_ms":1079,"p99_latency_ms":1158,"time_to_first_token_ms":222,"total_runtime_ms":981,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0253,"confidence_interval":0.9373,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4959,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":946,"p95_latency_ms":1156,"p99_latency_ms":1240,"time_to_first_token_ms":226,"total_runtime_ms":1051,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0207,"confidence_interval":0.9379,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3531,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":869,"p95_latency_ms":1062,"p99_latency_ms":1140,"time_to_first_token_ms":212,"total_runtime_ms":966,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.015,"confidence_interval":0.9302,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6804,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":934,"p95_latency_ms":1142,"p99_latency_ms":1225,"time_to_first_token_ms":216,"total_runtime_ms":1038,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0156,"confidence_interval":0.9447,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9096,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":896,"p95_latency_ms":1095,"p99_latency_ms":1175,"time_to_first_token_ms":235,"total_runtime_ms":996,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0277,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4842,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":922,"p95_latency_ms":1127,"p99_latency_ms":1209,"time_to_first_token_ms":241,"total_runtime_ms":1025,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0169,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0328,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":858,"p95_latency_ms":1049,"p99_latency_ms":1125,"time_to_first_token_ms":239,"total_runtime_ms":953,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":81,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0299,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4177,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2154,"p95_latency_ms":2633,"p99_latency_ms":2824,"time_to_first_token_ms":383,"total_runtime_ms":2393,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.94,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2626,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2122,"p95_latency_ms":2593,"p99_latency_ms":2782,"time_to_first_token_ms":355,"total_runtime_ms":2357,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.018,"confidence_interval":0.9357,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0225,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2165,"p95_latency_ms":2647,"p99_latency_ms":2839,"time_to_first_token_ms":368,"total_runtime_ms":2406,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0281,"confidence_interval":0.9396,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1288,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2158,"p95_latency_ms":2637,"p99_latency_ms":2829,"time_to_first_token_ms":366,"total_runtime_ms":2398,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0181,"confidence_interval":0.9369,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.19,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2204,"p95_latency_ms":2694,"p99_latency_ms":2890,"time_to_first_token_ms":381,"total_runtime_ms":2449,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0276,"confidence_interval":0.9378,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1447,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2129,"p95_latency_ms":2602,"p99_latency_ms":2791,"time_to_first_token_ms":354,"total_runtime_ms":2365,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0208,"confidence_interval":0.9467,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6713,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2162,"p95_latency_ms":2642,"p99_latency_ms":2834,"time_to_first_token_ms":380,"total_runtime_ms":2402,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9496,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4205,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2122,"p95_latency_ms":2594,"p99_latency_ms":2783,"time_to_first_token_ms":374,"total_runtime_ms":2358,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0184,"confidence_interval":0.9416,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6903,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2202,"p95_latency_ms":2691,"p99_latency_ms":2886,"time_to_first_token_ms":369,"total_runtime_ms":2446,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.015,"confidence_interval":0.9356,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.493,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2140,"p95_latency_ms":2615,"p99_latency_ms":2806,"time_to_first_token_ms":367,"total_runtime_ms":2378,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":100,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0242,"confidence_interval":0.9454,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9878,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1283,"p95_latency_ms":1568,"p99_latency_ms":1682,"time_to_first_token_ms":315,"total_runtime_ms":1425,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0294,"confidence_interval":0.9385,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.402,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1358,"p95_latency_ms":1660,"p99_latency_ms":1781,"time_to_first_token_ms":318,"total_runtime_ms":1509,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.029,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8702,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1338,"p95_latency_ms":1635,"p99_latency_ms":1754,"time_to_first_token_ms":333,"total_runtime_ms":1487,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0236,"confidence_interval":0.9339,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7653,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1318,"p95_latency_ms":1611,"p99_latency_ms":1728,"time_to_first_token_ms":327,"total_runtime_ms":1464,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0209,"confidence_interval":0.9304,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4442,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1346,"p95_latency_ms":1646,"p99_latency_ms":1765,"time_to_first_token_ms":328,"total_runtime_ms":1496,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0134,"confidence_interval":0.9456,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4647,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1329,"p95_latency_ms":1625,"p99_latency_ms":1743,"time_to_first_token_ms":322,"total_runtime_ms":1477,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0151,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9759,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1297,"p95_latency_ms":1585,"p99_latency_ms":1700,"time_to_first_token_ms":319,"total_runtime_ms":1441,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0117,"confidence_interval":0.939,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1508,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1345,"p95_latency_ms":1644,"p99_latency_ms":1764,"time_to_first_token_ms":310,"total_runtime_ms":1495,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0145,"confidence_interval":0.9339,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.536,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1270,"p95_latency_ms":1552,"p99_latency_ms":1665,"time_to_first_token_ms":318,"total_runtime_ms":1411,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0268,"confidence_interval":0.9494,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4468,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1306,"p95_latency_ms":1597,"p99_latency_ms":1713,"time_to_first_token_ms":307,"total_runtime_ms":1452,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":100,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0236,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3876,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":457,"p95_latency_ms":559,"p99_latency_ms":599,"time_to_first_token_ms":148,"total_runtime_ms":508,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0279,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1535,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":518,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":154,"total_runtime_ms":576,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0152,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8187,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":136,"total_runtime_ms":504,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0199,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0518,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":544,"p95_latency_ms":665,"p99_latency_ms":713,"time_to_first_token_ms":154,"total_runtime_ms":604,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0274,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4663,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":522,"p95_latency_ms":637,"p99_latency_ms":684,"time_to_first_token_ms":121,"total_runtime_ms":579,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0199,"confidence_interval":0.9414,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4816,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":535,"p95_latency_ms":654,"p99_latency_ms":702,"time_to_first_token_ms":138,"total_runtime_ms":595,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0247,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7488,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":502,"p95_latency_ms":613,"p99_latency_ms":658,"time_to_first_token_ms":154,"total_runtime_ms":557,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0227,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8344,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":538,"p95_latency_ms":658,"p99_latency_ms":706,"time_to_first_token_ms":150,"total_runtime_ms":598,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0289,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2833,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":529,"p95_latency_ms":647,"p99_latency_ms":694,"time_to_first_token_ms":143,"total_runtime_ms":588,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0234,"confidence_interval":0.9434,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8552,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":499,"p95_latency_ms":610,"p99_latency_ms":655,"time_to_first_token_ms":129,"total_runtime_ms":555,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0295,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6166,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":110,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0242,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6136,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":312,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":99,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0281,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.671,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":336,"p95_latency_ms":410,"p99_latency_ms":440,"time_to_first_token_ms":83,"total_runtime_ms":373,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0119,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1529,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":295,"p95_latency_ms":361,"p99_latency_ms":387,"time_to_first_token_ms":83,"total_runtime_ms":328,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0273,"confidence_interval":0.9452,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9473,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":300,"p95_latency_ms":366,"p99_latency_ms":393,"time_to_first_token_ms":104,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0267,"confidence_interval":0.9317,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9115,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":334,"p95_latency_ms":409,"p99_latency_ms":438,"time_to_first_token_ms":80,"total_runtime_ms":371,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0283,"confidence_interval":0.9362,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7138,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":369,"p95_latency_ms":451,"p99_latency_ms":484,"time_to_first_token_ms":94,"total_runtime_ms":410,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.022,"confidence_interval":0.9445,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0207,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":340,"p95_latency_ms":416,"p99_latency_ms":446,"time_to_first_token_ms":88,"total_runtime_ms":378,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0228,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9705,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":351,"p95_latency_ms":429,"p99_latency_ms":460,"time_to_first_token_ms":109,"total_runtime_ms":390,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0102,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.233,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":323,"p95_latency_ms":395,"p99_latency_ms":423,"time_to_first_token_ms":103,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0166,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.313,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":620,"p95_latency_ms":758,"p99_latency_ms":813,"time_to_first_token_ms":166,"total_runtime_ms":689,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.024,"confidence_interval":0.946,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2743,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":662,"p95_latency_ms":809,"p99_latency_ms":868,"time_to_first_token_ms":155,"total_runtime_ms":736,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0172,"confidence_interval":0.9352,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.57,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":686,"p95_latency_ms":839,"p99_latency_ms":900,"time_to_first_token_ms":160,"total_runtime_ms":762,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0214,"confidence_interval":0.9327,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3432,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":867,"time_to_first_token_ms":169,"total_runtime_ms":734,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0266,"confidence_interval":0.9381,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.106,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":600,"p95_latency_ms":734,"p99_latency_ms":787,"time_to_first_token_ms":162,"total_runtime_ms":667,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0105,"confidence_interval":0.9323,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.817,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":614,"p95_latency_ms":750,"p99_latency_ms":805,"time_to_first_token_ms":154,"total_runtime_ms":682,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0176,"confidence_interval":0.9322,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7372,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":614,"p95_latency_ms":750,"p99_latency_ms":805,"time_to_first_token_ms":142,"total_runtime_ms":682,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0224,"confidence_interval":0.9333,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.435,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":638,"p95_latency_ms":780,"p99_latency_ms":837,"time_to_first_token_ms":170,"total_runtime_ms":709,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0264,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3284,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":620,"p95_latency_ms":757,"p99_latency_ms":812,"time_to_first_token_ms":150,"total_runtime_ms":688,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.018,"confidence_interval":0.936,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3821,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":673,"p95_latency_ms":823,"p99_latency_ms":883,"time_to_first_token_ms":148,"total_runtime_ms":748,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.9457,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3766,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":367,"p95_latency_ms":449,"p99_latency_ms":482,"time_to_first_token_ms":101,"total_runtime_ms":408,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.019,"confidence_interval":0.9498,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7724,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":443,"p95_latency_ms":542,"p99_latency_ms":581,"time_to_first_token_ms":105,"total_runtime_ms":493,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0103,"confidence_interval":0.9319,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8408,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":446,"p95_latency_ms":546,"p99_latency_ms":585,"time_to_first_token_ms":98,"total_runtime_ms":496,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0129,"confidence_interval":0.9304,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7378,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":379,"p95_latency_ms":464,"p99_latency_ms":497,"time_to_first_token_ms":103,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0174,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2173,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":509,"time_to_first_token_ms":106,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0124,"confidence_interval":0.9414,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4587,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":373,"p95_latency_ms":456,"p99_latency_ms":490,"time_to_first_token_ms":119,"total_runtime_ms":415,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0151,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9169,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":449,"p95_latency_ms":549,"p99_latency_ms":589,"time_to_first_token_ms":110,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.9403,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9992,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":116,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0202,"confidence_interval":0.9367,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.848,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":357,"p95_latency_ms":436,"p99_latency_ms":468,"time_to_first_token_ms":121,"total_runtime_ms":396,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0106,"confidence_interval":0.9485,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6337,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":434,"p95_latency_ms":531,"p99_latency_ms":569,"time_to_first_token_ms":98,"total_runtime_ms":482,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0135,"confidence_interval":0.9434,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.956,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1249,"p95_latency_ms":1527,"p99_latency_ms":1638,"time_to_first_token_ms":226,"total_runtime_ms":1388,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.029,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1807,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1222,"p95_latency_ms":1493,"p99_latency_ms":1602,"time_to_first_token_ms":231,"total_runtime_ms":1358,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0113,"confidence_interval":0.9305,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5012,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1278,"p95_latency_ms":1562,"p99_latency_ms":1675,"time_to_first_token_ms":245,"total_runtime_ms":1420,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0163,"confidence_interval":0.9464,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5787,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1275,"p95_latency_ms":1559,"p99_latency_ms":1672,"time_to_first_token_ms":228,"total_runtime_ms":1417,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0111,"confidence_interval":0.9395,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2769,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1209,"p95_latency_ms":1477,"p99_latency_ms":1585,"time_to_first_token_ms":250,"total_runtime_ms":1343,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0261,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.706,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1223,"p95_latency_ms":1495,"p99_latency_ms":1604,"time_to_first_token_ms":244,"total_runtime_ms":1359,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0213,"confidence_interval":0.935,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.97,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1241,"p95_latency_ms":1517,"p99_latency_ms":1628,"time_to_first_token_ms":231,"total_runtime_ms":1379,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0109,"confidence_interval":0.9426,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.318,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1250,"p95_latency_ms":1527,"p99_latency_ms":1638,"time_to_first_token_ms":241,"total_runtime_ms":1389,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0196,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7188,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1196,"p95_latency_ms":1462,"p99_latency_ms":1568,"time_to_first_token_ms":238,"total_runtime_ms":1329,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0106,"confidence_interval":0.935,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.198,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1265,"p95_latency_ms":1546,"p99_latency_ms":1658,"time_to_first_token_ms":256,"total_runtime_ms":1405,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0117,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4204,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":730,"p95_latency_ms":892,"p99_latency_ms":957,"time_to_first_token_ms":199,"total_runtime_ms":811,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0171,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7707,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":779,"p95_latency_ms":953,"p99_latency_ms":1022,"time_to_first_token_ms":182,"total_runtime_ms":866,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0154,"confidence_interval":0.9369,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8306,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":720,"p95_latency_ms":880,"p99_latency_ms":944,"time_to_first_token_ms":181,"total_runtime_ms":800,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0285,"confidence_interval":0.9356,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9017,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":762,"p95_latency_ms":932,"p99_latency_ms":999,"time_to_first_token_ms":198,"total_runtime_ms":847,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0263,"confidence_interval":0.9349,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9637,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":773,"p95_latency_ms":945,"p99_latency_ms":1014,"time_to_first_token_ms":186,"total_runtime_ms":859,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0157,"confidence_interval":0.9445,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8095,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":737,"p95_latency_ms":900,"p99_latency_ms":966,"time_to_first_token_ms":198,"total_runtime_ms":818,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.027,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7436,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":776,"p95_latency_ms":948,"p99_latency_ms":1017,"time_to_first_token_ms":199,"total_runtime_ms":862,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0294,"confidence_interval":0.9306,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4802,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1009,"time_to_first_token_ms":184,"total_runtime_ms":855,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0215,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0256,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":739,"p95_latency_ms":904,"p99_latency_ms":969,"time_to_first_token_ms":207,"total_runtime_ms":821,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0136,"confidence_interval":0.9404,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5005,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:7b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":803,"p95_latency_ms":982,"p99_latency_ms":1053,"time_to_first_token_ms":186,"total_runtime_ms":892,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0285,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3996,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1834,"p95_latency_ms":2241,"p99_latency_ms":2404,"time_to_first_token_ms":334,"total_runtime_ms":2037,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.019,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9154,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1820,"p95_latency_ms":2225,"p99_latency_ms":2386,"time_to_first_token_ms":323,"total_runtime_ms":2022,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0197,"confidence_interval":0.9302,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0711,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1790,"p95_latency_ms":2187,"p99_latency_ms":2346,"time_to_first_token_ms":330,"total_runtime_ms":1988,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0237,"confidence_interval":0.9492,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1409,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1788,"p95_latency_ms":2186,"p99_latency_ms":2345,"time_to_first_token_ms":334,"total_runtime_ms":1987,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9216,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1822,"p95_latency_ms":2227,"p99_latency_ms":2389,"time_to_first_token_ms":319,"total_runtime_ms":2024,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0129,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1161,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1839,"p95_latency_ms":2248,"p99_latency_ms":2411,"time_to_first_token_ms":332,"total_runtime_ms":2044,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0198,"confidence_interval":0.9475,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8589,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1848,"p95_latency_ms":2259,"p99_latency_ms":2423,"time_to_first_token_ms":332,"total_runtime_ms":2054,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0144,"confidence_interval":0.9419,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6469,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1844,"p95_latency_ms":2254,"p99_latency_ms":2418,"time_to_first_token_ms":334,"total_runtime_ms":2049,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0184,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6799,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1863,"p95_latency_ms":2277,"p99_latency_ms":2442,"time_to_first_token_ms":309,"total_runtime_ms":2070,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0125,"confidence_interval":0.9429,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1117,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1775,"p95_latency_ms":2169,"p99_latency_ms":2327,"time_to_first_token_ms":327,"total_runtime_ms":1972,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0165,"confidence_interval":0.948,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5311,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1113,"p95_latency_ms":1360,"p99_latency_ms":1459,"time_to_first_token_ms":275,"total_runtime_ms":1237,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9428,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0229,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1089,"p95_latency_ms":1331,"p99_latency_ms":1428,"time_to_first_token_ms":285,"total_runtime_ms":1210,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0211,"confidence_interval":0.9492,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7604,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1124,"p95_latency_ms":1374,"p99_latency_ms":1474,"time_to_first_token_ms":265,"total_runtime_ms":1249,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0226,"confidence_interval":0.9384,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2133,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1135,"p95_latency_ms":1387,"p99_latency_ms":1488,"time_to_first_token_ms":269,"total_runtime_ms":1261,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0166,"confidence_interval":0.9369,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4556,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1090,"p95_latency_ms":1332,"p99_latency_ms":1429,"time_to_first_token_ms":287,"total_runtime_ms":1211,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0261,"confidence_interval":0.9417,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5884,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1066,"p95_latency_ms":1303,"p99_latency_ms":1397,"time_to_first_token_ms":258,"total_runtime_ms":1184,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0126,"confidence_interval":0.9346,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2493,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1150,"p95_latency_ms":1406,"p99_latency_ms":1508,"time_to_first_token_ms":273,"total_runtime_ms":1278,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0148,"confidence_interval":0.9464,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0619,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1093,"p95_latency_ms":1336,"p99_latency_ms":1433,"time_to_first_token_ms":278,"total_runtime_ms":1215,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0297,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8566,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1055,"p95_latency_ms":1290,"p99_latency_ms":1384,"time_to_first_token_ms":276,"total_runtime_ms":1173,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.023,"confidence_interval":0.9403,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8758,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"compare_structured|high|qwen2:14b|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1125,"p95_latency_ms":1375,"p99_latency_ms":1475,"time_to_first_token_ms":288,"total_runtime_ms":1250,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0268,"confidence_interval":0.9436,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5579,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":787,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":207,"total_runtime_ms":874,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0242,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2216,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":789,"p95_latency_ms":964,"p99_latency_ms":1034,"time_to_first_token_ms":207,"total_runtime_ms":876,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0167,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5959,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":832,"p95_latency_ms":1017,"p99_latency_ms":1091,"time_to_first_token_ms":196,"total_runtime_ms":925,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.029,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9603,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":842,"p95_latency_ms":1029,"p99_latency_ms":1103,"time_to_first_token_ms":211,"total_runtime_ms":935,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0121,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0457,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":750,"p95_latency_ms":916,"p99_latency_ms":983,"time_to_first_token_ms":200,"total_runtime_ms":833,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0164,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1678,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":826,"p95_latency_ms":1010,"p99_latency_ms":1083,"time_to_first_token_ms":195,"total_runtime_ms":918,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8653,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":759,"p95_latency_ms":927,"p99_latency_ms":995,"time_to_first_token_ms":222,"total_runtime_ms":843,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0133,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5357,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":773,"p95_latency_ms":945,"p99_latency_ms":1013,"time_to_first_token_ms":205,"total_runtime_ms":859,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0103,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5034,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":754,"p95_latency_ms":922,"p99_latency_ms":989,"time_to_first_token_ms":214,"total_runtime_ms":838,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0277,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1782,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1053,"time_to_first_token_ms":216,"total_runtime_ms":892,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0123,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9666,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":425,"p95_latency_ms":520,"p99_latency_ms":557,"time_to_first_token_ms":157,"total_runtime_ms":472,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.015,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4364,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":412,"p95_latency_ms":503,"p99_latency_ms":540,"time_to_first_token_ms":149,"total_runtime_ms":457,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0271,"confidence_interval":0.9397,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3141,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":397,"p95_latency_ms":485,"p99_latency_ms":521,"time_to_first_token_ms":156,"total_runtime_ms":441,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0247,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4314,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":400,"p95_latency_ms":489,"p99_latency_ms":524,"time_to_first_token_ms":145,"total_runtime_ms":444,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0188,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8594,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":467,"p95_latency_ms":571,"p99_latency_ms":613,"time_to_first_token_ms":155,"total_runtime_ms":519,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0146,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3956,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":155,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0142,"confidence_interval":0.9369,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3058,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":177,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0266,"confidence_interval":0.9428,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6958,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":393,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":166,"total_runtime_ms":437,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0227,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7315,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":408,"p95_latency_ms":498,"p99_latency_ms":535,"time_to_first_token_ms":147,"total_runtime_ms":453,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.027,"confidence_interval":0.9428,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6304,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":437,"p95_latency_ms":534,"p99_latency_ms":573,"time_to_first_token_ms":160,"total_runtime_ms":486,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0186,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5173,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1071,"p95_latency_ms":1309,"p99_latency_ms":1404,"time_to_first_token_ms":247,"total_runtime_ms":1190,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0231,"confidence_interval":0.9369,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7401,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1070,"p95_latency_ms":1308,"p99_latency_ms":1403,"time_to_first_token_ms":243,"total_runtime_ms":1189,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0148,"confidence_interval":0.9342,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2955,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1042,"p95_latency_ms":1274,"p99_latency_ms":1367,"time_to_first_token_ms":245,"total_runtime_ms":1158,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0184,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1198,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1111,"p95_latency_ms":1358,"p99_latency_ms":1457,"time_to_first_token_ms":241,"total_runtime_ms":1235,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0288,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0977,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1063,"p95_latency_ms":1299,"p99_latency_ms":1394,"time_to_first_token_ms":253,"total_runtime_ms":1181,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0149,"confidence_interval":0.948,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2568,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1119,"p95_latency_ms":1367,"p99_latency_ms":1467,"time_to_first_token_ms":262,"total_runtime_ms":1243,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0245,"confidence_interval":0.9308,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7057,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1097,"p95_latency_ms":1341,"p99_latency_ms":1438,"time_to_first_token_ms":253,"total_runtime_ms":1219,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0127,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6825,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1053,"p95_latency_ms":1287,"p99_latency_ms":1381,"time_to_first_token_ms":255,"total_runtime_ms":1170,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0149,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3173,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1051,"p95_latency_ms":1284,"p99_latency_ms":1378,"time_to_first_token_ms":257,"total_runtime_ms":1168,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0154,"confidence_interval":0.9419,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0344,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1069,"p95_latency_ms":1307,"p99_latency_ms":1402,"time_to_first_token_ms":235,"total_runtime_ms":1188,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0123,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3808,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":586,"p95_latency_ms":716,"p99_latency_ms":768,"time_to_first_token_ms":217,"total_runtime_ms":651,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0297,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7134,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":591,"p95_latency_ms":722,"p99_latency_ms":775,"time_to_first_token_ms":215,"total_runtime_ms":657,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0193,"confidence_interval":0.9437,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4153,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":555,"p95_latency_ms":679,"p99_latency_ms":728,"time_to_first_token_ms":210,"total_runtime_ms":617,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0291,"confidence_interval":0.94,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3134,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":560,"p95_latency_ms":684,"p99_latency_ms":734,"time_to_first_token_ms":214,"total_runtime_ms":622,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0237,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9793,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":581,"p95_latency_ms":710,"p99_latency_ms":761,"time_to_first_token_ms":206,"total_runtime_ms":645,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0192,"confidence_interval":0.9397,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4311,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":196,"total_runtime_ms":650,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0182,"confidence_interval":0.934,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3119,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":516,"p95_latency_ms":630,"p99_latency_ms":676,"time_to_first_token_ms":203,"total_runtime_ms":573,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0218,"confidence_interval":0.9347,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0821,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":547,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":218,"total_runtime_ms":608,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0233,"confidence_interval":0.9496,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.384,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":540,"p95_latency_ms":660,"p99_latency_ms":708,"time_to_first_token_ms":213,"total_runtime_ms":600,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9406,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5352,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":546,"p95_latency_ms":667,"p99_latency_ms":716,"time_to_first_token_ms":191,"total_runtime_ms":606,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0264,"confidence_interval":0.9466,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.441,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2055,"p95_latency_ms":2512,"p99_latency_ms":2694,"time_to_first_token_ms":407,"total_runtime_ms":2283,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0198,"confidence_interval":0.9308,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9712,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2040,"p95_latency_ms":2493,"p99_latency_ms":2675,"time_to_first_token_ms":429,"total_runtime_ms":2267,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0202,"confidence_interval":0.9431,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5395,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2105,"p95_latency_ms":2573,"p99_latency_ms":2760,"time_to_first_token_ms":423,"total_runtime_ms":2339,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.025,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2464,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2068,"p95_latency_ms":2527,"p99_latency_ms":2711,"time_to_first_token_ms":440,"total_runtime_ms":2297,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0208,"confidence_interval":0.9452,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.939,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2127,"p95_latency_ms":2599,"p99_latency_ms":2788,"time_to_first_token_ms":416,"total_runtime_ms":2363,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0122,"confidence_interval":0.9487,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5874,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2042,"p95_latency_ms":2495,"p99_latency_ms":2677,"time_to_first_token_ms":413,"total_runtime_ms":2268,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0267,"confidence_interval":0.9435,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3753,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2046,"p95_latency_ms":2501,"p99_latency_ms":2683,"time_to_first_token_ms":426,"total_runtime_ms":2274,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6201,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2038,"p95_latency_ms":2491,"p99_latency_ms":2672,"time_to_first_token_ms":409,"total_runtime_ms":2265,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0275,"confidence_interval":0.9487,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0282,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2028,"p95_latency_ms":2478,"p99_latency_ms":2659,"time_to_first_token_ms":408,"total_runtime_ms":2253,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0263,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2044,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2061,"p95_latency_ms":2519,"p99_latency_ms":2702,"time_to_first_token_ms":429,"total_runtime_ms":2290,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0196,"confidence_interval":0.9329,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1485,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1090,"p95_latency_ms":1332,"p99_latency_ms":1429,"time_to_first_token_ms":386,"total_runtime_ms":1211,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0259,"confidence_interval":0.9325,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.239,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1033,"p95_latency_ms":1263,"p99_latency_ms":1355,"time_to_first_token_ms":375,"total_runtime_ms":1148,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.022,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7563,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1035,"p95_latency_ms":1266,"p99_latency_ms":1358,"time_to_first_token_ms":386,"total_runtime_ms":1151,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0287,"confidence_interval":0.9331,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7862,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1064,"p95_latency_ms":1301,"p99_latency_ms":1395,"time_to_first_token_ms":381,"total_runtime_ms":1182,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0125,"confidence_interval":0.9492,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9385,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1065,"p95_latency_ms":1302,"p99_latency_ms":1396,"time_to_first_token_ms":376,"total_runtime_ms":1183,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0225,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1486,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1035,"p95_latency_ms":1264,"p99_latency_ms":1356,"time_to_first_token_ms":382,"total_runtime_ms":1149,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0174,"confidence_interval":0.9497,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2838,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1057,"p95_latency_ms":1292,"p99_latency_ms":1386,"time_to_first_token_ms":372,"total_runtime_ms":1175,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0141,"confidence_interval":0.9316,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.926,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1047,"p95_latency_ms":1280,"p99_latency_ms":1373,"time_to_first_token_ms":375,"total_runtime_ms":1163,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0235,"confidence_interval":0.9395,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4384,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1071,"p95_latency_ms":1309,"p99_latency_ms":1404,"time_to_first_token_ms":392,"total_runtime_ms":1190,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0153,"confidence_interval":0.9384,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3564,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:7b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1344,"time_to_first_token_ms":396,"total_runtime_ms":1139,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0189,"confidence_interval":0.9427,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9663,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3097,"p95_latency_ms":3785,"p99_latency_ms":4060,"time_to_first_token_ms":592,"total_runtime_ms":3441,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0128,"confidence_interval":0.9302,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1806,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3058,"p95_latency_ms":3738,"p99_latency_ms":4010,"time_to_first_token_ms":584,"total_runtime_ms":3398,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.029,"confidence_interval":0.9478,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7955,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3034,"p95_latency_ms":3709,"p99_latency_ms":3979,"time_to_first_token_ms":575,"total_runtime_ms":3372,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0299,"confidence_interval":0.9357,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7135,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3085,"p95_latency_ms":3771,"p99_latency_ms":4045,"time_to_first_token_ms":582,"total_runtime_ms":3428,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0199,"confidence_interval":0.9432,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3819,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3048,"p95_latency_ms":3726,"p99_latency_ms":3997,"time_to_first_token_ms":585,"total_runtime_ms":3387,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0149,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0785,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3059,"p95_latency_ms":3738,"p99_latency_ms":4010,"time_to_first_token_ms":599,"total_runtime_ms":3399,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0281,"confidence_interval":0.9424,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9021,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3091,"p95_latency_ms":3778,"p99_latency_ms":4053,"time_to_first_token_ms":575,"total_runtime_ms":3434,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.943,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3429,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3053,"p95_latency_ms":3731,"p99_latency_ms":4002,"time_to_first_token_ms":604,"total_runtime_ms":3392,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0197,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9536,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3018,"p95_latency_ms":3689,"p99_latency_ms":3957,"time_to_first_token_ms":586,"total_runtime_ms":3354,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0145,"confidence_interval":0.9349,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8447,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3089,"p95_latency_ms":3776,"p99_latency_ms":4050,"time_to_first_token_ms":606,"total_runtime_ms":3432,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0291,"confidence_interval":0.9466,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9392,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1511,"p95_latency_ms":1847,"p99_latency_ms":1981,"time_to_first_token_ms":554,"total_runtime_ms":1679,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0251,"confidence_interval":0.9458,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6287,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1550,"p95_latency_ms":1895,"p99_latency_ms":2033,"time_to_first_token_ms":547,"total_runtime_ms":1723,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0155,"confidence_interval":0.9489,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6145,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1541,"p95_latency_ms":1883,"p99_latency_ms":2020,"time_to_first_token_ms":537,"total_runtime_ms":1712,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0101,"confidence_interval":0.9498,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4364,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1575,"p95_latency_ms":1925,"p99_latency_ms":2065,"time_to_first_token_ms":531,"total_runtime_ms":1750,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.9457,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.418,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1517,"p95_latency_ms":1854,"p99_latency_ms":1989,"time_to_first_token_ms":548,"total_runtime_ms":1685,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0103,"confidence_interval":0.9382,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1872,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1524,"p95_latency_ms":1863,"p99_latency_ms":1998,"time_to_first_token_ms":549,"total_runtime_ms":1694,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0273,"confidence_interval":0.9452,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6739,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1587,"p95_latency_ms":1939,"p99_latency_ms":2080,"time_to_first_token_ms":556,"total_runtime_ms":1763,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0127,"confidence_interval":0.9344,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8952,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1559,"p95_latency_ms":1906,"p99_latency_ms":2044,"time_to_first_token_ms":558,"total_runtime_ms":1732,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0291,"confidence_interval":0.9411,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5992,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1598,"p95_latency_ms":1953,"p99_latency_ms":2095,"time_to_first_token_ms":528,"total_runtime_ms":1776,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0138,"confidence_interval":0.9396,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3158,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|low|qwen2:14b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1515,"p95_latency_ms":1852,"p99_latency_ms":1986,"time_to_first_token_ms":552,"total_runtime_ms":1683,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.01,"confidence_interval":0.9379,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8428,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":575,"p95_latency_ms":702,"p99_latency_ms":753,"time_to_first_token_ms":191,"total_runtime_ms":638,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0291,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7897,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":583,"p95_latency_ms":712,"p99_latency_ms":764,"time_to_first_token_ms":177,"total_runtime_ms":647,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.027,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.178,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":592,"p95_latency_ms":723,"p99_latency_ms":776,"time_to_first_token_ms":168,"total_runtime_ms":657,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0209,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6561,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":652,"p95_latency_ms":797,"p99_latency_ms":855,"time_to_first_token_ms":161,"total_runtime_ms":724,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0218,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5634,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":669,"p95_latency_ms":817,"p99_latency_ms":877,"time_to_first_token_ms":164,"total_runtime_ms":743,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.014,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7812,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":161,"total_runtime_ms":635,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0239,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0206,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":581,"p95_latency_ms":711,"p99_latency_ms":762,"time_to_first_token_ms":158,"total_runtime_ms":646,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0115,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9078,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":867,"time_to_first_token_ms":160,"total_runtime_ms":735,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0183,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2087,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":639,"p95_latency_ms":781,"p99_latency_ms":838,"time_to_first_token_ms":162,"total_runtime_ms":710,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0182,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8169,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":639,"p95_latency_ms":781,"p99_latency_ms":838,"time_to_first_token_ms":187,"total_runtime_ms":710,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0245,"confidence_interval":0.9395,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9914,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":377,"p95_latency_ms":460,"p99_latency_ms":494,"time_to_first_token_ms":135,"total_runtime_ms":419,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.593,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":298,"p95_latency_ms":364,"p99_latency_ms":390,"time_to_first_token_ms":140,"total_runtime_ms":331,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.017,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1388,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":323,"p95_latency_ms":395,"p99_latency_ms":424,"time_to_first_token_ms":123,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.022,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1714,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":287,"p95_latency_ms":351,"p99_latency_ms":377,"time_to_first_token_ms":133,"total_runtime_ms":319,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.016,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9109,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":461,"p99_latency_ms":495,"time_to_first_token_ms":139,"total_runtime_ms":419,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0251,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3982,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":388,"p95_latency_ms":474,"p99_latency_ms":508,"time_to_first_token_ms":129,"total_runtime_ms":431,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0203,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8721,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":407,"time_to_first_token_ms":129,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0275,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4382,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":355,"p95_latency_ms":434,"p99_latency_ms":466,"time_to_first_token_ms":125,"total_runtime_ms":395,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.014,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6611,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":332,"p95_latency_ms":405,"p99_latency_ms":435,"time_to_first_token_ms":121,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0281,"confidence_interval":0.9376,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4137,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":377,"p99_latency_ms":404,"time_to_first_token_ms":119,"total_runtime_ms":342,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0185,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7771,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":860,"p95_latency_ms":1051,"p99_latency_ms":1127,"time_to_first_token_ms":202,"total_runtime_ms":955,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0182,"confidence_interval":0.9441,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7042,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":766,"p95_latency_ms":936,"p99_latency_ms":1004,"time_to_first_token_ms":200,"total_runtime_ms":851,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.016,"confidence_interval":0.9316,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0718,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":780,"p95_latency_ms":954,"p99_latency_ms":1023,"time_to_first_token_ms":208,"total_runtime_ms":867,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0287,"confidence_interval":0.9435,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3679,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":845,"p95_latency_ms":1032,"p99_latency_ms":1107,"time_to_first_token_ms":207,"total_runtime_ms":938,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0105,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1224,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":779,"p95_latency_ms":952,"p99_latency_ms":1021,"time_to_first_token_ms":191,"total_runtime_ms":865,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0156,"confidence_interval":0.9402,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6464,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":862,"p95_latency_ms":1054,"p99_latency_ms":1130,"time_to_first_token_ms":217,"total_runtime_ms":958,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0112,"confidence_interval":0.9434,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4268,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":797,"p95_latency_ms":974,"p99_latency_ms":1045,"time_to_first_token_ms":210,"total_runtime_ms":885,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9469,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6228,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":813,"p95_latency_ms":993,"p99_latency_ms":1066,"time_to_first_token_ms":196,"total_runtime_ms":903,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0167,"confidence_interval":0.9444,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0718,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":764,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":217,"total_runtime_ms":848,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.018,"confidence_interval":0.9372,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1965,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":861,"p95_latency_ms":1053,"p99_latency_ms":1129,"time_to_first_token_ms":210,"total_runtime_ms":957,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0208,"confidence_interval":0.9413,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6621,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":167,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0164,"confidence_interval":0.9334,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.88,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":402,"p95_latency_ms":492,"p99_latency_ms":527,"time_to_first_token_ms":151,"total_runtime_ms":447,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0534,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":631,"time_to_first_token_ms":157,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0224,"confidence_interval":0.9351,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.52,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":452,"p95_latency_ms":552,"p99_latency_ms":592,"time_to_first_token_ms":178,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0201,"confidence_interval":0.9379,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7573,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":577,"time_to_first_token_ms":153,"total_runtime_ms":489,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0104,"confidence_interval":0.9313,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1993,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":179,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0186,"confidence_interval":0.9411,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5459,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":384,"p95_latency_ms":469,"p99_latency_ms":503,"time_to_first_token_ms":145,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0135,"confidence_interval":0.945,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.228,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":409,"p95_latency_ms":500,"p99_latency_ms":536,"time_to_first_token_ms":170,"total_runtime_ms":455,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0294,"confidence_interval":0.931,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9847,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":150,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0102,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5226,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":512,"time_to_first_token_ms":152,"total_runtime_ms":434,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0168,"confidence_interval":0.9467,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6722,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1618,"p95_latency_ms":1977,"p99_latency_ms":2121,"time_to_first_token_ms":348,"total_runtime_ms":1797,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0154,"confidence_interval":0.949,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2656,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1606,"p95_latency_ms":1962,"p99_latency_ms":2105,"time_to_first_token_ms":323,"total_runtime_ms":1784,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0142,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7386,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1556,"p95_latency_ms":1902,"p99_latency_ms":2040,"time_to_first_token_ms":321,"total_runtime_ms":1729,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0206,"confidence_interval":0.9409,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1655,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1622,"p95_latency_ms":1983,"p99_latency_ms":2127,"time_to_first_token_ms":321,"total_runtime_ms":1802,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0277,"confidence_interval":0.9363,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2562,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1614,"p95_latency_ms":1972,"p99_latency_ms":2116,"time_to_first_token_ms":334,"total_runtime_ms":1793,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0107,"confidence_interval":0.9443,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0924,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1563,"p95_latency_ms":1911,"p99_latency_ms":2050,"time_to_first_token_ms":346,"total_runtime_ms":1737,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0149,"confidence_interval":0.9477,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4564,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1538,"p95_latency_ms":1880,"p99_latency_ms":2016,"time_to_first_token_ms":338,"total_runtime_ms":1709,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0282,"confidence_interval":0.9392,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2877,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1528,"p95_latency_ms":1868,"p99_latency_ms":2004,"time_to_first_token_ms":334,"total_runtime_ms":1698,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0192,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3691,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1605,"p95_latency_ms":1962,"p99_latency_ms":2105,"time_to_first_token_ms":320,"total_runtime_ms":1784,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0123,"confidence_interval":0.9313,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7168,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1530,"p95_latency_ms":1870,"p99_latency_ms":2006,"time_to_first_token_ms":320,"total_runtime_ms":1700,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0207,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9474,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":812,"p95_latency_ms":993,"p99_latency_ms":1065,"time_to_first_token_ms":302,"total_runtime_ms":903,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0231,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6075,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":795,"p95_latency_ms":971,"p99_latency_ms":1042,"time_to_first_token_ms":291,"total_runtime_ms":883,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0145,"confidence_interval":0.9306,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7212,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":813,"p95_latency_ms":994,"p99_latency_ms":1066,"time_to_first_token_ms":285,"total_runtime_ms":904,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0232,"confidence_interval":0.9399,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6116,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":844,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":276,"total_runtime_ms":938,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0122,"confidence_interval":0.9499,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4592,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":819,"p95_latency_ms":1001,"p99_latency_ms":1074,"time_to_first_token_ms":284,"total_runtime_ms":910,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0195,"confidence_interval":0.9313,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.969,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":814,"p95_latency_ms":995,"p99_latency_ms":1068,"time_to_first_token_ms":297,"total_runtime_ms":905,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0288,"confidence_interval":0.9349,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.565,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":802,"p95_latency_ms":980,"p99_latency_ms":1052,"time_to_first_token_ms":307,"total_runtime_ms":891,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0108,"confidence_interval":0.9462,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5831,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":761,"p95_latency_ms":931,"p99_latency_ms":998,"time_to_first_token_ms":295,"total_runtime_ms":846,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0183,"confidence_interval":0.9308,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.988,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":858,"p95_latency_ms":1048,"p99_latency_ms":1124,"time_to_first_token_ms":299,"total_runtime_ms":953,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0232,"confidence_interval":0.9443,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3998,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":844,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":297,"total_runtime_ms":937,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0248,"confidence_interval":0.9391,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0468,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2293,"p95_latency_ms":2802,"p99_latency_ms":3006,"time_to_first_token_ms":448,"total_runtime_ms":2547,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.024,"confidence_interval":0.9385,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4607,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2300,"p95_latency_ms":2811,"p99_latency_ms":3016,"time_to_first_token_ms":461,"total_runtime_ms":2556,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0111,"confidence_interval":0.9364,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6635,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2249,"p95_latency_ms":2749,"p99_latency_ms":2948,"time_to_first_token_ms":476,"total_runtime_ms":2499,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0233,"confidence_interval":0.9411,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3728,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2271,"p95_latency_ms":2776,"p99_latency_ms":2978,"time_to_first_token_ms":465,"total_runtime_ms":2523,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0294,"confidence_interval":0.932,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.855,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2295,"p95_latency_ms":2805,"p99_latency_ms":3009,"time_to_first_token_ms":456,"total_runtime_ms":2550,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0259,"confidence_interval":0.9323,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4971,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2310,"p95_latency_ms":2823,"p99_latency_ms":3028,"time_to_first_token_ms":458,"total_runtime_ms":2566,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0169,"confidence_interval":0.9378,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3836,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2352,"p95_latency_ms":2874,"p99_latency_ms":3083,"time_to_first_token_ms":454,"total_runtime_ms":2613,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9388,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1371,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2295,"p95_latency_ms":2805,"p99_latency_ms":3009,"time_to_first_token_ms":446,"total_runtime_ms":2550,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0163,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9571,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2290,"p95_latency_ms":2799,"p99_latency_ms":3003,"time_to_first_token_ms":471,"total_runtime_ms":2545,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0194,"confidence_interval":0.9491,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9024,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2289,"p95_latency_ms":2798,"p99_latency_ms":3002,"time_to_first_token_ms":467,"total_runtime_ms":2544,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0255,"confidence_interval":0.9456,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7722,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1140,"p95_latency_ms":1394,"p99_latency_ms":1495,"time_to_first_token_ms":405,"total_runtime_ms":1267,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0123,"confidence_interval":0.9364,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.407,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1169,"p95_latency_ms":1429,"p99_latency_ms":1533,"time_to_first_token_ms":413,"total_runtime_ms":1299,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0108,"confidence_interval":0.9445,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9867,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1184,"p95_latency_ms":1447,"p99_latency_ms":1553,"time_to_first_token_ms":403,"total_runtime_ms":1316,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0285,"confidence_interval":0.9304,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9012,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1202,"p95_latency_ms":1470,"p99_latency_ms":1576,"time_to_first_token_ms":424,"total_runtime_ms":1336,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.013,"confidence_interval":0.9315,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6253,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1168,"p95_latency_ms":1428,"p99_latency_ms":1532,"time_to_first_token_ms":403,"total_runtime_ms":1298,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0195,"confidence_interval":0.9404,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3891,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1136,"p95_latency_ms":1388,"p99_latency_ms":1489,"time_to_first_token_ms":434,"total_runtime_ms":1262,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0133,"confidence_interval":0.9393,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8135,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1140,"p95_latency_ms":1393,"p99_latency_ms":1495,"time_to_first_token_ms":405,"total_runtime_ms":1267,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0149,"confidence_interval":0.9475,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3632,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1201,"p95_latency_ms":1468,"p99_latency_ms":1575,"time_to_first_token_ms":410,"total_runtime_ms":1335,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0106,"confidence_interval":0.9492,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4175,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1171,"p95_latency_ms":1432,"p99_latency_ms":1536,"time_to_first_token_ms":412,"total_runtime_ms":1301,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0103,"confidence_interval":0.9375,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.232,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1137,"p95_latency_ms":1389,"p99_latency_ms":1490,"time_to_first_token_ms":415,"total_runtime_ms":1263,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0165,"confidence_interval":0.9411,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6374,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":498,"p95_latency_ms":609,"p99_latency_ms":653,"time_to_first_token_ms":165,"total_runtime_ms":553,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0211,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7806,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":520,"p95_latency_ms":635,"p99_latency_ms":682,"time_to_first_token_ms":163,"total_runtime_ms":578,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0249,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4077,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":465,"p95_latency_ms":569,"p99_latency_ms":610,"time_to_first_token_ms":141,"total_runtime_ms":517,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0178,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2799,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":672,"time_to_first_token_ms":164,"total_runtime_ms":569,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0262,"confidence_interval":0.9329,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5255,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":483,"p95_latency_ms":591,"p99_latency_ms":634,"time_to_first_token_ms":147,"total_runtime_ms":537,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0296,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5726,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":529,"p95_latency_ms":647,"p99_latency_ms":694,"time_to_first_token_ms":136,"total_runtime_ms":588,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0113,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2497,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":134,"total_runtime_ms":535,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0159,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5915,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":523,"p95_latency_ms":639,"p99_latency_ms":686,"time_to_first_token_ms":142,"total_runtime_ms":581,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0256,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7281,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":437,"p95_latency_ms":534,"p99_latency_ms":573,"time_to_first_token_ms":151,"total_runtime_ms":486,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0159,"confidence_interval":0.9419,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9495,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":554,"time_to_first_token_ms":132,"total_runtime_ms":470,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0975,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":232,"p95_latency_ms":283,"p99_latency_ms":304,"time_to_first_token_ms":98,"total_runtime_ms":257,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0167,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4357,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":230,"p95_latency_ms":281,"p99_latency_ms":301,"time_to_first_token_ms":109,"total_runtime_ms":255,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0146,"confidence_interval":0.9452,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1005,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":315,"p95_latency_ms":385,"p99_latency_ms":413,"time_to_first_token_ms":88,"total_runtime_ms":350,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0191,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5908,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":260,"p95_latency_ms":318,"p99_latency_ms":341,"time_to_first_token_ms":89,"total_runtime_ms":289,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0292,"confidence_interval":0.9327,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2924,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":235,"p95_latency_ms":288,"p99_latency_ms":309,"time_to_first_token_ms":118,"total_runtime_ms":262,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0288,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1399,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":408,"time_to_first_token_ms":94,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0139,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0856,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":234,"p95_latency_ms":286,"p99_latency_ms":306,"time_to_first_token_ms":94,"total_runtime_ms":260,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0237,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3145,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":223,"p95_latency_ms":273,"p99_latency_ms":293,"time_to_first_token_ms":110,"total_runtime_ms":248,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.025,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.077,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":230,"p95_latency_ms":281,"p99_latency_ms":302,"time_to_first_token_ms":118,"total_runtime_ms":256,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0236,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4818,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":291,"p99_latency_ms":313,"time_to_first_token_ms":96,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0293,"confidence_interval":0.9378,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3342,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":625,"p95_latency_ms":763,"p99_latency_ms":819,"time_to_first_token_ms":186,"total_runtime_ms":694,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9428,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6065,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":606,"p95_latency_ms":741,"p99_latency_ms":795,"time_to_first_token_ms":184,"total_runtime_ms":674,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0222,"confidence_interval":0.9318,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5462,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":607,"p95_latency_ms":742,"p99_latency_ms":796,"time_to_first_token_ms":170,"total_runtime_ms":675,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0159,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4285,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":593,"p95_latency_ms":725,"p99_latency_ms":778,"time_to_first_token_ms":159,"total_runtime_ms":659,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0132,"confidence_interval":0.9325,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3001,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":676,"p95_latency_ms":826,"p99_latency_ms":886,"time_to_first_token_ms":176,"total_runtime_ms":751,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0164,"confidence_interval":0.9334,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.785,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":618,"p95_latency_ms":755,"p99_latency_ms":810,"time_to_first_token_ms":176,"total_runtime_ms":687,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0139,"confidence_interval":0.9385,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2869,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":636,"p95_latency_ms":778,"p99_latency_ms":834,"time_to_first_token_ms":182,"total_runtime_ms":707,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0238,"confidence_interval":0.9314,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3452,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":867,"time_to_first_token_ms":158,"total_runtime_ms":735,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0168,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8851,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":640,"p95_latency_ms":782,"p99_latency_ms":838,"time_to_first_token_ms":185,"total_runtime_ms":711,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0281,"confidence_interval":0.9467,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4602,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":780,"time_to_first_token_ms":186,"total_runtime_ms":661,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0159,"confidence_interval":0.9404,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0367,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":354,"p95_latency_ms":433,"p99_latency_ms":464,"time_to_first_token_ms":137,"total_runtime_ms":393,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0229,"confidence_interval":0.9307,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8956,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":337,"p95_latency_ms":412,"p99_latency_ms":442,"time_to_first_token_ms":140,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0264,"confidence_interval":0.9483,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6793,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":477,"time_to_first_token_ms":132,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0125,"confidence_interval":0.9442,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0035,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":371,"p95_latency_ms":453,"p99_latency_ms":486,"time_to_first_token_ms":128,"total_runtime_ms":412,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0181,"confidence_interval":0.943,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0027,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":294,"p95_latency_ms":360,"p99_latency_ms":386,"time_to_first_token_ms":119,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0258,"confidence_interval":0.9337,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9399,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":476,"p99_latency_ms":510,"time_to_first_token_ms":117,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0281,"confidence_interval":0.9322,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2221,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":353,"p95_latency_ms":431,"p99_latency_ms":463,"time_to_first_token_ms":146,"total_runtime_ms":392,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0225,"confidence_interval":0.9444,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4418,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":298,"p95_latency_ms":364,"p99_latency_ms":391,"time_to_first_token_ms":118,"total_runtime_ms":331,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0193,"confidence_interval":0.9342,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5163,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":366,"p95_latency_ms":447,"p99_latency_ms":480,"time_to_first_token_ms":125,"total_runtime_ms":407,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.02,"confidence_interval":0.9353,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2259,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":382,"p95_latency_ms":467,"p99_latency_ms":501,"time_to_first_token_ms":137,"total_runtime_ms":425,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0283,"confidence_interval":0.9426,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1325,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1238,"p95_latency_ms":1513,"p99_latency_ms":1623,"time_to_first_token_ms":288,"total_runtime_ms":1376,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0279,"confidence_interval":0.9418,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2468,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1199,"p95_latency_ms":1466,"p99_latency_ms":1572,"time_to_first_token_ms":259,"total_runtime_ms":1333,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0112,"confidence_interval":0.9497,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3457,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1164,"p95_latency_ms":1423,"p99_latency_ms":1526,"time_to_first_token_ms":262,"total_runtime_ms":1294,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0118,"confidence_interval":0.9433,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1594,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1227,"p95_latency_ms":1500,"p99_latency_ms":1609,"time_to_first_token_ms":273,"total_runtime_ms":1363,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0262,"confidence_interval":0.9446,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1568,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1139,"p95_latency_ms":1392,"p99_latency_ms":1493,"time_to_first_token_ms":256,"total_runtime_ms":1266,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0201,"confidence_interval":0.9449,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4321,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1211,"p95_latency_ms":1481,"p99_latency_ms":1588,"time_to_first_token_ms":272,"total_runtime_ms":1346,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0143,"confidence_interval":0.933,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7736,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1139,"p95_latency_ms":1392,"p99_latency_ms":1493,"time_to_first_token_ms":275,"total_runtime_ms":1265,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0139,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9135,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1184,"p95_latency_ms":1447,"p99_latency_ms":1553,"time_to_first_token_ms":258,"total_runtime_ms":1316,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0135,"confidence_interval":0.9325,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6453,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1230,"p95_latency_ms":1503,"p99_latency_ms":1612,"time_to_first_token_ms":267,"total_runtime_ms":1366,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0146,"confidence_interval":0.9366,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.656,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1173,"p95_latency_ms":1433,"p99_latency_ms":1538,"time_to_first_token_ms":262,"total_runtime_ms":1303,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0116,"confidence_interval":0.9322,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4783,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":594,"p95_latency_ms":726,"p99_latency_ms":779,"time_to_first_token_ms":235,"total_runtime_ms":660,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.013,"confidence_interval":0.9332,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4274,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":586,"p95_latency_ms":716,"p99_latency_ms":768,"time_to_first_token_ms":235,"total_runtime_ms":651,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0229,"confidence_interval":0.947,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.3183,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":575,"p95_latency_ms":703,"p99_latency_ms":754,"time_to_first_token_ms":240,"total_runtime_ms":639,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0123,"confidence_interval":0.9404,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1736,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":602,"p95_latency_ms":736,"p99_latency_ms":789,"time_to_first_token_ms":238,"total_runtime_ms":669,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.026,"confidence_interval":0.941,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5241,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":611,"p95_latency_ms":747,"p99_latency_ms":801,"time_to_first_token_ms":232,"total_runtime_ms":679,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0249,"confidence_interval":0.9487,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9555,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":638,"p95_latency_ms":780,"p99_latency_ms":837,"time_to_first_token_ms":221,"total_runtime_ms":709,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0189,"confidence_interval":0.9384,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6964,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":669,"p95_latency_ms":818,"p99_latency_ms":878,"time_to_first_token_ms":218,"total_runtime_ms":744,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0246,"confidence_interval":0.9329,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3331,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":604,"p95_latency_ms":739,"p99_latency_ms":792,"time_to_first_token_ms":212,"total_runtime_ms":671,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0129,"confidence_interval":0.9435,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2292,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":604,"p95_latency_ms":738,"p99_latency_ms":792,"time_to_first_token_ms":231,"total_runtime_ms":671,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0271,"confidence_interval":0.939,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8866,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":663,"p95_latency_ms":810,"p99_latency_ms":869,"time_to_first_token_ms":219,"total_runtime_ms":737,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0258,"confidence_interval":0.9367,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0187,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1754,"p95_latency_ms":2143,"p99_latency_ms":2299,"time_to_first_token_ms":382,"total_runtime_ms":1949,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0121,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9014,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1685,"p95_latency_ms":2060,"p99_latency_ms":2210,"time_to_first_token_ms":353,"total_runtime_ms":1873,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0236,"confidence_interval":0.94,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1265,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1711,"p95_latency_ms":2092,"p99_latency_ms":2244,"time_to_first_token_ms":373,"total_runtime_ms":1902,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0185,"confidence_interval":0.9415,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7684,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1694,"p95_latency_ms":2071,"p99_latency_ms":2221,"time_to_first_token_ms":351,"total_runtime_ms":1882,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0212,"confidence_interval":0.9484,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1374,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1752,"p95_latency_ms":2141,"p99_latency_ms":2297,"time_to_first_token_ms":377,"total_runtime_ms":1946,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0147,"confidence_interval":0.9496,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9011,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1693,"p95_latency_ms":2070,"p99_latency_ms":2220,"time_to_first_token_ms":358,"total_runtime_ms":1882,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0115,"confidence_interval":0.9415,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8988,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1753,"p95_latency_ms":2142,"p99_latency_ms":2298,"time_to_first_token_ms":376,"total_runtime_ms":1948,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0127,"confidence_interval":0.9483,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4351,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1701,"p95_latency_ms":2079,"p99_latency_ms":2230,"time_to_first_token_ms":369,"total_runtime_ms":1890,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0267,"confidence_interval":0.9451,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6606,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1769,"p95_latency_ms":2162,"p99_latency_ms":2319,"time_to_first_token_ms":382,"total_runtime_ms":1965,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0155,"confidence_interval":0.9457,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4232,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1769,"p95_latency_ms":2162,"p99_latency_ms":2319,"time_to_first_token_ms":359,"total_runtime_ms":1965,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.024,"confidence_interval":0.9322,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3806,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":923,"p95_latency_ms":1128,"p99_latency_ms":1210,"time_to_first_token_ms":334,"total_runtime_ms":1025,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.9463,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1005,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":923,"p95_latency_ms":1129,"p99_latency_ms":1211,"time_to_first_token_ms":334,"total_runtime_ms":1026,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0143,"confidence_interval":0.9303,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.697,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":882,"p95_latency_ms":1078,"p99_latency_ms":1157,"time_to_first_token_ms":334,"total_runtime_ms":980,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0203,"confidence_interval":0.9443,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6501,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":865,"p95_latency_ms":1057,"p99_latency_ms":1134,"time_to_first_token_ms":333,"total_runtime_ms":961,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0138,"confidence_interval":0.9416,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7668,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1030,"p99_latency_ms":1105,"time_to_first_token_ms":322,"total_runtime_ms":937,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0201,"confidence_interval":0.9479,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1327,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":849,"p95_latency_ms":1037,"p99_latency_ms":1113,"time_to_first_token_ms":321,"total_runtime_ms":943,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0143,"confidence_interval":0.9455,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7456,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":872,"p95_latency_ms":1065,"p99_latency_ms":1143,"time_to_first_token_ms":321,"total_runtime_ms":968,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0117,"confidence_interval":0.9354,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9635,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":934,"p95_latency_ms":1141,"p99_latency_ms":1224,"time_to_first_token_ms":338,"total_runtime_ms":1038,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0127,"confidence_interval":0.9498,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9312,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":945,"p95_latency_ms":1155,"p99_latency_ms":1239,"time_to_first_token_ms":321,"total_runtime_ms":1050,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.9403,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.496,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":856,"p95_latency_ms":1046,"p99_latency_ms":1122,"time_to_first_token_ms":331,"total_runtime_ms":951,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0167,"confidence_interval":0.9317,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9097,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":374,"p95_latency_ms":457,"p99_latency_ms":490,"time_to_first_token_ms":131,"total_runtime_ms":415,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0176,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.282,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":357,"p95_latency_ms":436,"p99_latency_ms":467,"time_to_first_token_ms":136,"total_runtime_ms":396,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5912,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":152,"total_runtime_ms":488,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0264,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6502,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":452,"p95_latency_ms":552,"p99_latency_ms":592,"time_to_first_token_ms":144,"total_runtime_ms":502,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0188,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.629,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":443,"p95_latency_ms":542,"p99_latency_ms":581,"time_to_first_token_ms":121,"total_runtime_ms":493,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0225,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8031,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":447,"p95_latency_ms":546,"p99_latency_ms":585,"time_to_first_token_ms":136,"total_runtime_ms":496,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0155,"confidence_interval":0.9317,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0339,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":437,"p95_latency_ms":534,"p99_latency_ms":573,"time_to_first_token_ms":148,"total_runtime_ms":485,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0151,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3611,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":483,"time_to_first_token_ms":123,"total_runtime_ms":409,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6853,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":419,"p95_latency_ms":512,"p99_latency_ms":549,"time_to_first_token_ms":122,"total_runtime_ms":465,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0163,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4111,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":427,"p95_latency_ms":522,"p99_latency_ms":560,"time_to_first_token_ms":135,"total_runtime_ms":474,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0147,"confidence_interval":0.9423,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0477,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":249,"p95_latency_ms":304,"p99_latency_ms":327,"time_to_first_token_ms":104,"total_runtime_ms":277,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0255,"confidence_interval":0.9466,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9907,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":211,"p95_latency_ms":258,"p99_latency_ms":277,"time_to_first_token_ms":107,"total_runtime_ms":235,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0187,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9964,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":190,"p95_latency_ms":232,"p99_latency_ms":249,"time_to_first_token_ms":89,"total_runtime_ms":211,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0112,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6909,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":204,"p95_latency_ms":250,"p99_latency_ms":268,"time_to_first_token_ms":79,"total_runtime_ms":227,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.021,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5175,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":251,"p95_latency_ms":307,"p99_latency_ms":329,"time_to_first_token_ms":78,"total_runtime_ms":279,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0139,"confidence_interval":0.9451,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3994,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":247,"p95_latency_ms":302,"p99_latency_ms":324,"time_to_first_token_ms":76,"total_runtime_ms":275,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0114,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9723,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":203,"p95_latency_ms":248,"p99_latency_ms":266,"time_to_first_token_ms":83,"total_runtime_ms":225,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0262,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6852,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":215,"p95_latency_ms":262,"p99_latency_ms":282,"time_to_first_token_ms":103,"total_runtime_ms":239,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.025,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3658,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":226,"p95_latency_ms":276,"p99_latency_ms":296,"time_to_first_token_ms":100,"total_runtime_ms":251,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0268,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3592,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":189,"p95_latency_ms":231,"p99_latency_ms":247,"time_to_first_token_ms":87,"total_runtime_ms":210,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.023,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8902,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":749,"time_to_first_token_ms":159,"total_runtime_ms":635,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0256,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1983,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":528,"p95_latency_ms":645,"p99_latency_ms":692,"time_to_first_token_ms":159,"total_runtime_ms":587,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.027,"confidence_interval":0.9393,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1684,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":744,"time_to_first_token_ms":164,"total_runtime_ms":630,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0101,"confidence_interval":0.9381,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8775,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":498,"p95_latency_ms":608,"p99_latency_ms":652,"time_to_first_token_ms":156,"total_runtime_ms":553,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0223,"confidence_interval":0.9337,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9611,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":503,"p95_latency_ms":615,"p99_latency_ms":660,"time_to_first_token_ms":165,"total_runtime_ms":559,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0211,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0965,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":630,"time_to_first_token_ms":171,"total_runtime_ms":534,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0226,"confidence_interval":0.9458,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6836,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":522,"p95_latency_ms":638,"p99_latency_ms":685,"time_to_first_token_ms":147,"total_runtime_ms":580,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.012,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9421,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":554,"p95_latency_ms":678,"p99_latency_ms":727,"time_to_first_token_ms":162,"total_runtime_ms":616,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0206,"confidence_interval":0.9317,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0173,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":505,"p95_latency_ms":618,"p99_latency_ms":663,"time_to_first_token_ms":143,"total_runtime_ms":562,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0207,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2112,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":165,"total_runtime_ms":535,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0207,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2435,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":300,"p95_latency_ms":367,"p99_latency_ms":394,"time_to_first_token_ms":114,"total_runtime_ms":334,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0175,"confidence_interval":0.9373,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1061,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":319,"p95_latency_ms":390,"p99_latency_ms":418,"time_to_first_token_ms":107,"total_runtime_ms":354,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0236,"confidence_interval":0.9485,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3586,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":289,"p95_latency_ms":354,"p99_latency_ms":380,"time_to_first_token_ms":113,"total_runtime_ms":322,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.02,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3781,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":321,"p95_latency_ms":393,"p99_latency_ms":421,"time_to_first_token_ms":96,"total_runtime_ms":357,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0175,"confidence_interval":0.9479,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8758,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":340,"p95_latency_ms":416,"p99_latency_ms":446,"time_to_first_token_ms":99,"total_runtime_ms":378,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0273,"confidence_interval":0.9362,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4523,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":291,"p95_latency_ms":355,"p99_latency_ms":381,"time_to_first_token_ms":112,"total_runtime_ms":323,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.015,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2258,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":290,"p95_latency_ms":355,"p99_latency_ms":381,"time_to_first_token_ms":96,"total_runtime_ms":323,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0256,"confidence_interval":0.948,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9552,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":254,"p95_latency_ms":310,"p99_latency_ms":333,"time_to_first_token_ms":126,"total_runtime_ms":282,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0142,"confidence_interval":0.9459,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.198,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":291,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":113,"total_runtime_ms":323,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0137,"confidence_interval":0.9457,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7644,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":317,"p95_latency_ms":388,"p99_latency_ms":416,"time_to_first_token_ms":117,"total_runtime_ms":352,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0163,"confidence_interval":0.9462,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3323,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":956,"p95_latency_ms":1168,"p99_latency_ms":1253,"time_to_first_token_ms":241,"total_runtime_ms":1062,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0137,"confidence_interval":0.9362,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2464,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1023,"p95_latency_ms":1250,"p99_latency_ms":1341,"time_to_first_token_ms":237,"total_runtime_ms":1137,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.013,"confidence_interval":0.9411,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0013,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":970,"p95_latency_ms":1186,"p99_latency_ms":1272,"time_to_first_token_ms":240,"total_runtime_ms":1078,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0176,"confidence_interval":0.9313,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0788,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1047,"p95_latency_ms":1280,"p99_latency_ms":1373,"time_to_first_token_ms":255,"total_runtime_ms":1164,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.017,"confidence_interval":0.9309,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9894,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":989,"p95_latency_ms":1208,"p99_latency_ms":1296,"time_to_first_token_ms":256,"total_runtime_ms":1099,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0119,"confidence_interval":0.9312,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.061,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1015,"p95_latency_ms":1240,"p99_latency_ms":1330,"time_to_first_token_ms":253,"total_runtime_ms":1127,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0129,"confidence_interval":0.9497,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.726,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":962,"p95_latency_ms":1175,"p99_latency_ms":1261,"time_to_first_token_ms":227,"total_runtime_ms":1068,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0235,"confidence_interval":0.9326,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2948,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":967,"p95_latency_ms":1181,"p99_latency_ms":1267,"time_to_first_token_ms":255,"total_runtime_ms":1074,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0241,"confidence_interval":0.9462,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4911,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":963,"p95_latency_ms":1178,"p99_latency_ms":1263,"time_to_first_token_ms":254,"total_runtime_ms":1070,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0205,"confidence_interval":0.9486,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6696,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":977,"p95_latency_ms":1194,"p99_latency_ms":1281,"time_to_first_token_ms":231,"total_runtime_ms":1085,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0204,"confidence_interval":0.9444,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3637,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":671,"time_to_first_token_ms":202,"total_runtime_ms":569,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.021,"confidence_interval":0.9416,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9292,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":563,"p95_latency_ms":688,"p99_latency_ms":738,"time_to_first_token_ms":180,"total_runtime_ms":625,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0123,"confidence_interval":0.9321,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0895,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":560,"p95_latency_ms":685,"p99_latency_ms":735,"time_to_first_token_ms":197,"total_runtime_ms":623,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0284,"confidence_interval":0.9461,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1204,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":528,"p95_latency_ms":645,"p99_latency_ms":692,"time_to_first_token_ms":185,"total_runtime_ms":587,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.012,"confidence_interval":0.9318,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.456,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":515,"p95_latency_ms":630,"p99_latency_ms":676,"time_to_first_token_ms":183,"total_runtime_ms":573,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0175,"confidence_interval":0.9376,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1768,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":538,"p95_latency_ms":657,"p99_latency_ms":705,"time_to_first_token_ms":178,"total_runtime_ms":597,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0272,"confidence_interval":0.9398,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8681,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":545,"p95_latency_ms":666,"p99_latency_ms":714,"time_to_first_token_ms":200,"total_runtime_ms":605,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0101,"confidence_interval":0.9382,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0863,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":543,"p95_latency_ms":664,"p99_latency_ms":712,"time_to_first_token_ms":209,"total_runtime_ms":603,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0283,"confidence_interval":0.931,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6802,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":543,"p95_latency_ms":663,"p99_latency_ms":711,"time_to_first_token_ms":189,"total_runtime_ms":603,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0243,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.769,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:7b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":190,"total_runtime_ms":591,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.018,"confidence_interval":0.9469,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5497,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1467,"p95_latency_ms":1793,"p99_latency_ms":1923,"time_to_first_token_ms":311,"total_runtime_ms":1630,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0293,"confidence_interval":0.9465,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8223,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1487,"p95_latency_ms":1817,"p99_latency_ms":1949,"time_to_first_token_ms":314,"total_runtime_ms":1652,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0247,"confidence_interval":0.946,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1328,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1487,"p95_latency_ms":1817,"p99_latency_ms":1950,"time_to_first_token_ms":304,"total_runtime_ms":1652,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0285,"confidence_interval":0.9349,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5179,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1447,"p95_latency_ms":1768,"p99_latency_ms":1897,"time_to_first_token_ms":305,"total_runtime_ms":1607,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0175,"confidence_interval":0.932,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9541,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1419,"p95_latency_ms":1734,"p99_latency_ms":1860,"time_to_first_token_ms":312,"total_runtime_ms":1577,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0133,"confidence_interval":0.9484,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1649,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1488,"p95_latency_ms":1818,"p99_latency_ms":1950,"time_to_first_token_ms":327,"total_runtime_ms":1653,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0146,"confidence_interval":0.9355,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4686,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1411,"p95_latency_ms":1725,"p99_latency_ms":1850,"time_to_first_token_ms":334,"total_runtime_ms":1568,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0201,"confidence_interval":0.9301,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1831,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1475,"p95_latency_ms":1803,"p99_latency_ms":1934,"time_to_first_token_ms":318,"total_runtime_ms":1639,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0147,"confidence_interval":0.9304,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5669,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1407,"p95_latency_ms":1719,"p99_latency_ms":1844,"time_to_first_token_ms":314,"total_runtime_ms":1563,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.082,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1479,"p95_latency_ms":1807,"p99_latency_ms":1939,"time_to_first_token_ms":319,"total_runtime_ms":1643,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0118,"confidence_interval":0.9406,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9665,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":766,"p95_latency_ms":937,"p99_latency_ms":1005,"time_to_first_token_ms":263,"total_runtime_ms":852,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0287,"confidence_interval":0.945,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7125,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":712,"p95_latency_ms":870,"p99_latency_ms":933,"time_to_first_token_ms":287,"total_runtime_ms":791,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0113,"confidence_interval":0.9419,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3697,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1003,"time_to_first_token_ms":267,"total_runtime_ms":850,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.012,"confidence_interval":0.9323,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1307,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":748,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":274,"total_runtime_ms":831,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0266,"confidence_interval":0.946,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9437,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":748,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":285,"total_runtime_ms":831,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.027,"confidence_interval":0.9356,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1937,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":744,"p95_latency_ms":909,"p99_latency_ms":976,"time_to_first_token_ms":258,"total_runtime_ms":827,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0272,"confidence_interval":0.9362,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7189,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":780,"p95_latency_ms":953,"p99_latency_ms":1023,"time_to_first_token_ms":281,"total_runtime_ms":867,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0292,"confidence_interval":0.9418,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5459,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":773,"p95_latency_ms":945,"p99_latency_ms":1013,"time_to_first_token_ms":275,"total_runtime_ms":859,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0241,"confidence_interval":0.9399,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4087,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":772,"p95_latency_ms":944,"p99_latency_ms":1012,"time_to_first_token_ms":281,"total_runtime_ms":858,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0263,"confidence_interval":0.9435,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5432,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"ambiguous_intent|high|qwen2:14b|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":785,"p95_latency_ms":959,"p99_latency_ms":1029,"time_to_first_token_ms":266,"total_runtime_ms":872,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0103,"confidence_interval":0.9351,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9176,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":907,"p95_latency_ms":1108,"p99_latency_ms":1189,"time_to_first_token_ms":193,"total_runtime_ms":1007,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.99,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":915,"p95_latency_ms":1118,"p99_latency_ms":1200,"time_to_first_token_ms":197,"total_runtime_ms":1017,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0226,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3139,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":954,"p95_latency_ms":1166,"p99_latency_ms":1250,"time_to_first_token_ms":209,"total_runtime_ms":1060,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0136,"confidence_interval":0.9445,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7688,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":895,"p95_latency_ms":1093,"p99_latency_ms":1173,"time_to_first_token_ms":197,"total_runtime_ms":994,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9208,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":929,"p95_latency_ms":1136,"p99_latency_ms":1218,"time_to_first_token_ms":220,"total_runtime_ms":1032,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0257,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.821,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":898,"p95_latency_ms":1097,"p99_latency_ms":1177,"time_to_first_token_ms":191,"total_runtime_ms":997,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0161,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1335,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":963,"p95_latency_ms":1177,"p99_latency_ms":1262,"time_to_first_token_ms":204,"total_runtime_ms":1070,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0116,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6765,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":963,"p95_latency_ms":1177,"p99_latency_ms":1262,"time_to_first_token_ms":209,"total_runtime_ms":1070,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0249,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3472,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":906,"p95_latency_ms":1107,"p99_latency_ms":1188,"time_to_first_token_ms":211,"total_runtime_ms":1007,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0246,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9175,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":908,"p95_latency_ms":1109,"p99_latency_ms":1190,"time_to_first_token_ms":198,"total_runtime_ms":1009,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0142,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6663,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":151,"total_runtime_ms":650,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0286,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5938,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":547,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":174,"total_runtime_ms":608,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0216,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5446,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":557,"p95_latency_ms":681,"p99_latency_ms":730,"time_to_first_token_ms":159,"total_runtime_ms":619,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0209,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9058,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":525,"p95_latency_ms":641,"p99_latency_ms":688,"time_to_first_token_ms":144,"total_runtime_ms":583,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0188,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3558,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":785,"time_to_first_token_ms":149,"total_runtime_ms":665,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6472,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":589,"p95_latency_ms":719,"p99_latency_ms":772,"time_to_first_token_ms":146,"total_runtime_ms":654,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.012,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1828,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":600,"p95_latency_ms":733,"p99_latency_ms":787,"time_to_first_token_ms":153,"total_runtime_ms":667,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.9351,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4485,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":536,"p95_latency_ms":656,"p99_latency_ms":703,"time_to_first_token_ms":175,"total_runtime_ms":596,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.027,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0615,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":570,"p95_latency_ms":697,"p99_latency_ms":748,"time_to_first_token_ms":165,"total_runtime_ms":634,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0217,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6829,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":531,"p95_latency_ms":649,"p99_latency_ms":696,"time_to_first_token_ms":144,"total_runtime_ms":590,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0205,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0647,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1229,"p95_latency_ms":1502,"p99_latency_ms":1611,"time_to_first_token_ms":241,"total_runtime_ms":1365,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0235,"confidence_interval":0.9327,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6991,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1218,"p95_latency_ms":1488,"p99_latency_ms":1597,"time_to_first_token_ms":252,"total_runtime_ms":1353,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0114,"confidence_interval":0.9353,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8989,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1226,"p95_latency_ms":1498,"p99_latency_ms":1607,"time_to_first_token_ms":246,"total_runtime_ms":1362,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0243,"confidence_interval":0.9418,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1812,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1282,"p95_latency_ms":1566,"p99_latency_ms":1680,"time_to_first_token_ms":248,"total_runtime_ms":1424,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0227,"confidence_interval":0.9316,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8689,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1240,"p95_latency_ms":1516,"p99_latency_ms":1626,"time_to_first_token_ms":253,"total_runtime_ms":1378,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0278,"confidence_interval":0.9378,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6231,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1260,"p95_latency_ms":1540,"p99_latency_ms":1652,"time_to_first_token_ms":248,"total_runtime_ms":1400,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0158,"confidence_interval":0.9494,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6691,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1243,"p95_latency_ms":1519,"p99_latency_ms":1629,"time_to_first_token_ms":238,"total_runtime_ms":1381,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.018,"confidence_interval":0.937,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0314,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1201,"p95_latency_ms":1468,"p99_latency_ms":1574,"time_to_first_token_ms":235,"total_runtime_ms":1334,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.026,"confidence_interval":0.9476,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.215,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1201,"p95_latency_ms":1468,"p99_latency_ms":1574,"time_to_first_token_ms":242,"total_runtime_ms":1334,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0177,"confidence_interval":0.9437,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8001,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1231,"p95_latency_ms":1505,"p99_latency_ms":1614,"time_to_first_token_ms":243,"total_runtime_ms":1368,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0267,"confidence_interval":0.9397,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3582,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":726,"p95_latency_ms":888,"p99_latency_ms":952,"time_to_first_token_ms":213,"total_runtime_ms":807,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.026,"confidence_interval":0.937,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.356,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":195,"total_runtime_ms":765,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0289,"confidence_interval":0.9411,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.648,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":701,"p95_latency_ms":856,"p99_latency_ms":919,"time_to_first_token_ms":207,"total_runtime_ms":778,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0198,"confidence_interval":0.9354,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.684,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":762,"p95_latency_ms":931,"p99_latency_ms":999,"time_to_first_token_ms":222,"total_runtime_ms":847,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0135,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6345,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":766,"p95_latency_ms":936,"p99_latency_ms":1005,"time_to_first_token_ms":213,"total_runtime_ms":851,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0246,"confidence_interval":0.9331,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8889,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":786,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":208,"total_runtime_ms":873,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0191,"confidence_interval":0.9484,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2009,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":702,"p95_latency_ms":858,"p99_latency_ms":920,"time_to_first_token_ms":211,"total_runtime_ms":780,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0129,"confidence_interval":0.9491,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2025,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":787,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":219,"total_runtime_ms":874,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0159,"confidence_interval":0.947,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5334,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":751,"p95_latency_ms":918,"p99_latency_ms":985,"time_to_first_token_ms":220,"total_runtime_ms":835,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0191,"confidence_interval":0.9432,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9223,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":194,"total_runtime_ms":765,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0161,"confidence_interval":0.9462,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7214,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2454,"p95_latency_ms":3000,"p99_latency_ms":3218,"time_to_first_token_ms":433,"total_runtime_ms":2727,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0184,"confidence_interval":0.9323,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3396,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2427,"p95_latency_ms":2967,"p99_latency_ms":3182,"time_to_first_token_ms":425,"total_runtime_ms":2697,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0268,"confidence_interval":0.9462,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2892,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2432,"p95_latency_ms":2973,"p99_latency_ms":3189,"time_to_first_token_ms":412,"total_runtime_ms":2703,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0228,"confidence_interval":0.9452,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7673,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2423,"p95_latency_ms":2961,"p99_latency_ms":3177,"time_to_first_token_ms":426,"total_runtime_ms":2692,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0271,"confidence_interval":0.9412,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8571,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2483,"p95_latency_ms":3035,"p99_latency_ms":3256,"time_to_first_token_ms":427,"total_runtime_ms":2759,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0128,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.481,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2400,"p95_latency_ms":2933,"p99_latency_ms":3147,"time_to_first_token_ms":435,"total_runtime_ms":2667,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0159,"confidence_interval":0.948,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0408,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2458,"p95_latency_ms":3005,"p99_latency_ms":3223,"time_to_first_token_ms":440,"total_runtime_ms":2731,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0201,"confidence_interval":0.9399,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2475,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2410,"p95_latency_ms":2946,"p99_latency_ms":3160,"time_to_first_token_ms":430,"total_runtime_ms":2678,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0188,"confidence_interval":0.9408,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4324,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2473,"p95_latency_ms":3022,"p99_latency_ms":3242,"time_to_first_token_ms":439,"total_runtime_ms":2747,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0222,"confidence_interval":0.9321,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3722,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2422,"p95_latency_ms":2960,"p99_latency_ms":3175,"time_to_first_token_ms":406,"total_runtime_ms":2691,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0264,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1706,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1394,"p95_latency_ms":1704,"p99_latency_ms":1828,"time_to_first_token_ms":394,"total_runtime_ms":1549,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.024,"confidence_interval":0.9363,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8076,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1367,"p95_latency_ms":1670,"p99_latency_ms":1792,"time_to_first_token_ms":380,"total_runtime_ms":1518,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0117,"confidence_interval":0.9392,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2428,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1402,"p95_latency_ms":1714,"p99_latency_ms":1838,"time_to_first_token_ms":377,"total_runtime_ms":1558,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0296,"confidence_interval":0.9489,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1818,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1413,"p95_latency_ms":1727,"p99_latency_ms":1853,"time_to_first_token_ms":376,"total_runtime_ms":1570,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0288,"confidence_interval":0.9329,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2321,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1375,"p95_latency_ms":1681,"p99_latency_ms":1803,"time_to_first_token_ms":368,"total_runtime_ms":1528,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0126,"confidence_interval":0.9447,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9868,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1389,"p95_latency_ms":1698,"p99_latency_ms":1822,"time_to_first_token_ms":393,"total_runtime_ms":1544,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0293,"confidence_interval":0.9363,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6412,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1381,"p95_latency_ms":1688,"p99_latency_ms":1811,"time_to_first_token_ms":362,"total_runtime_ms":1535,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0268,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.8488,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1453,"p95_latency_ms":1776,"p99_latency_ms":1906,"time_to_first_token_ms":386,"total_runtime_ms":1615,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0207,"confidence_interval":0.9463,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3854,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1377,"p95_latency_ms":1683,"p99_latency_ms":1805,"time_to_first_token_ms":389,"total_runtime_ms":1530,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0139,"confidence_interval":0.9353,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.042,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:7b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1421,"p95_latency_ms":1737,"p99_latency_ms":1863,"time_to_first_token_ms":392,"total_runtime_ms":1579,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.023,"confidence_interval":0.9346,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7203,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3562,"p95_latency_ms":4354,"p99_latency_ms":4671,"time_to_first_token_ms":586,"total_runtime_ms":3958,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0192,"confidence_interval":0.9357,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0925,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3612,"p95_latency_ms":4415,"p99_latency_ms":4736,"time_to_first_token_ms":602,"total_runtime_ms":4013,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0114,"confidence_interval":0.9324,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1569,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3543,"p95_latency_ms":4330,"p99_latency_ms":4645,"time_to_first_token_ms":586,"total_runtime_ms":3936,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0172,"confidence_interval":0.9411,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0156,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3572,"p95_latency_ms":4366,"p99_latency_ms":4683,"time_to_first_token_ms":582,"total_runtime_ms":3969,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0148,"confidence_interval":0.9341,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3511,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3562,"p95_latency_ms":4353,"p99_latency_ms":4670,"time_to_first_token_ms":583,"total_runtime_ms":3958,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0193,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8253,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3578,"p95_latency_ms":4373,"p99_latency_ms":4691,"time_to_first_token_ms":580,"total_runtime_ms":3975,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0143,"confidence_interval":0.9494,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8608,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3589,"p95_latency_ms":4387,"p99_latency_ms":4706,"time_to_first_token_ms":601,"total_runtime_ms":3988,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0188,"confidence_interval":0.9365,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1106,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3579,"p95_latency_ms":4374,"p99_latency_ms":4692,"time_to_first_token_ms":603,"total_runtime_ms":3976,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0168,"confidence_interval":0.935,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0132,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3593,"p95_latency_ms":4391,"p99_latency_ms":4711,"time_to_first_token_ms":590,"total_runtime_ms":3992,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.014,"confidence_interval":0.9483,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8145,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3626,"p95_latency_ms":4432,"p99_latency_ms":4754,"time_to_first_token_ms":606,"total_runtime_ms":4029,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0297,"confidence_interval":0.9322,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.969,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2026,"p95_latency_ms":2477,"p99_latency_ms":2657,"time_to_first_token_ms":536,"total_runtime_ms":2251,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0259,"confidence_interval":0.9468,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6029,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2109,"p95_latency_ms":2578,"p99_latency_ms":2765,"time_to_first_token_ms":550,"total_runtime_ms":2343,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0216,"confidence_interval":0.9353,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.835,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2106,"p95_latency_ms":2574,"p99_latency_ms":2761,"time_to_first_token_ms":560,"total_runtime_ms":2340,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0237,"confidence_interval":0.9455,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1906,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2046,"p95_latency_ms":2501,"p99_latency_ms":2683,"time_to_first_token_ms":528,"total_runtime_ms":2274,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0142,"confidence_interval":0.9391,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3517,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2073,"p95_latency_ms":2533,"p99_latency_ms":2717,"time_to_first_token_ms":530,"total_runtime_ms":2303,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0167,"confidence_interval":0.9445,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.078,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2094,"p95_latency_ms":2559,"p99_latency_ms":2745,"time_to_first_token_ms":556,"total_runtime_ms":2326,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0142,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5599,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2088,"p95_latency_ms":2552,"p99_latency_ms":2738,"time_to_first_token_ms":551,"total_runtime_ms":2320,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0207,"confidence_interval":0.9464,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9081,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2046,"p95_latency_ms":2501,"p99_latency_ms":2683,"time_to_first_token_ms":540,"total_runtime_ms":2273,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0202,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0097,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2045,"p95_latency_ms":2499,"p99_latency_ms":2681,"time_to_first_token_ms":556,"total_runtime_ms":2272,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.025,"confidence_interval":0.9424,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6637,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|low|qwen2:14b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2033,"p95_latency_ms":2485,"p99_latency_ms":2666,"time_to_first_token_ms":536,"total_runtime_ms":2259,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0123,"confidence_interval":0.9454,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8518,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":724,"p95_latency_ms":885,"p99_latency_ms":949,"time_to_first_token_ms":179,"total_runtime_ms":804,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0172,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.475,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":672,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":161,"total_runtime_ms":747,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0273,"confidence_interval":0.9479,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0251,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":704,"p95_latency_ms":860,"p99_latency_ms":923,"time_to_first_token_ms":173,"total_runtime_ms":782,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0228,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4203,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":663,"p95_latency_ms":810,"p99_latency_ms":869,"time_to_first_token_ms":167,"total_runtime_ms":736,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0288,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5362,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":712,"p95_latency_ms":870,"p99_latency_ms":933,"time_to_first_token_ms":176,"total_runtime_ms":791,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0111,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9438,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":696,"p95_latency_ms":851,"p99_latency_ms":913,"time_to_first_token_ms":169,"total_runtime_ms":773,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0187,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5382,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":709,"p95_latency_ms":867,"p99_latency_ms":930,"time_to_first_token_ms":163,"total_runtime_ms":788,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0154,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3607,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":698,"p95_latency_ms":853,"p99_latency_ms":915,"time_to_first_token_ms":175,"total_runtime_ms":776,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0115,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7714,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":728,"p95_latency_ms":890,"p99_latency_ms":954,"time_to_first_token_ms":157,"total_runtime_ms":809,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0264,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7539,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":685,"p95_latency_ms":837,"p99_latency_ms":898,"time_to_first_token_ms":182,"total_runtime_ms":761,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0232,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8275,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":421,"p95_latency_ms":515,"p99_latency_ms":552,"time_to_first_token_ms":127,"total_runtime_ms":468,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0108,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2083,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":443,"p95_latency_ms":541,"p99_latency_ms":580,"time_to_first_token_ms":137,"total_runtime_ms":492,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0255,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6305,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":426,"p95_latency_ms":521,"p99_latency_ms":559,"time_to_first_token_ms":121,"total_runtime_ms":474,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0271,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4478,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":507,"time_to_first_token_ms":139,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.754,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":460,"p95_latency_ms":562,"p99_latency_ms":603,"time_to_first_token_ms":135,"total_runtime_ms":511,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.013,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5912,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":513,"time_to_first_token_ms":132,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0154,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.612,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":412,"p95_latency_ms":503,"p99_latency_ms":540,"time_to_first_token_ms":137,"total_runtime_ms":457,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0266,"confidence_interval":0.942,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6323,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":601,"time_to_first_token_ms":123,"total_runtime_ms":510,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0295,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9777,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":431,"p95_latency_ms":527,"p99_latency_ms":565,"time_to_first_token_ms":117,"total_runtime_ms":479,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0172,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6111,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":464,"p95_latency_ms":567,"p99_latency_ms":608,"time_to_first_token_ms":126,"total_runtime_ms":515,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0164,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6729,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":894,"p95_latency_ms":1092,"p99_latency_ms":1172,"time_to_first_token_ms":204,"total_runtime_ms":993,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.011,"confidence_interval":0.933,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0963,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":974,"p95_latency_ms":1190,"p99_latency_ms":1277,"time_to_first_token_ms":206,"total_runtime_ms":1082,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0184,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4851,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":901,"p95_latency_ms":1101,"p99_latency_ms":1181,"time_to_first_token_ms":200,"total_runtime_ms":1001,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0134,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9564,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":948,"p95_latency_ms":1159,"p99_latency_ms":1243,"time_to_first_token_ms":195,"total_runtime_ms":1054,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0153,"confidence_interval":0.9379,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.428,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":981,"p95_latency_ms":1199,"p99_latency_ms":1286,"time_to_first_token_ms":215,"total_runtime_ms":1090,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0251,"confidence_interval":0.9433,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3552,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":962,"p95_latency_ms":1176,"p99_latency_ms":1261,"time_to_first_token_ms":206,"total_runtime_ms":1069,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0175,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0579,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":909,"p95_latency_ms":1111,"p99_latency_ms":1192,"time_to_first_token_ms":218,"total_runtime_ms":1010,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0248,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3059,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":965,"p95_latency_ms":1180,"p99_latency_ms":1265,"time_to_first_token_ms":221,"total_runtime_ms":1072,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.01,"confidence_interval":0.9341,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1625,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":920,"p95_latency_ms":1125,"p99_latency_ms":1206,"time_to_first_token_ms":198,"total_runtime_ms":1022,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0274,"confidence_interval":0.9445,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5041,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":947,"p95_latency_ms":1158,"p99_latency_ms":1242,"time_to_first_token_ms":223,"total_runtime_ms":1052,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0277,"confidence_interval":0.9448,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2911,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":785,"time_to_first_token_ms":154,"total_runtime_ms":665,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0155,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2896,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":569,"p95_latency_ms":695,"p99_latency_ms":746,"time_to_first_token_ms":175,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0119,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5921,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":528,"p95_latency_ms":646,"p99_latency_ms":693,"time_to_first_token_ms":168,"total_runtime_ms":587,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0115,"confidence_interval":0.9322,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4863,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":618,"p95_latency_ms":756,"p99_latency_ms":811,"time_to_first_token_ms":155,"total_runtime_ms":687,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0273,"confidence_interval":0.9444,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3904,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":780,"time_to_first_token_ms":173,"total_runtime_ms":661,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0133,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2074,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":530,"p95_latency_ms":648,"p99_latency_ms":695,"time_to_first_token_ms":163,"total_runtime_ms":589,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0113,"confidence_interval":0.9389,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2968,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":603,"p95_latency_ms":737,"p99_latency_ms":790,"time_to_first_token_ms":174,"total_runtime_ms":670,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.012,"confidence_interval":0.9346,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2197,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":610,"p95_latency_ms":745,"p99_latency_ms":799,"time_to_first_token_ms":162,"total_runtime_ms":677,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0176,"confidence_interval":0.9344,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3427,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":551,"p95_latency_ms":673,"p99_latency_ms":722,"time_to_first_token_ms":163,"total_runtime_ms":612,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0127,"confidence_interval":0.9495,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7267,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":536,"p95_latency_ms":656,"p99_latency_ms":703,"time_to_first_token_ms":176,"total_runtime_ms":596,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.027,"confidence_interval":0.9439,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4882,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1839,"p95_latency_ms":2248,"p99_latency_ms":2411,"time_to_first_token_ms":345,"total_runtime_ms":2044,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0283,"confidence_interval":0.9345,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3376,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1880,"p95_latency_ms":2298,"p99_latency_ms":2466,"time_to_first_token_ms":334,"total_runtime_ms":2089,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0271,"confidence_interval":0.9454,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7282,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1874,"p95_latency_ms":2290,"p99_latency_ms":2457,"time_to_first_token_ms":330,"total_runtime_ms":2082,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0196,"confidence_interval":0.9496,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2506,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1868,"p95_latency_ms":2283,"p99_latency_ms":2449,"time_to_first_token_ms":334,"total_runtime_ms":2076,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0103,"confidence_interval":0.949,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4518,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1836,"p95_latency_ms":2243,"p99_latency_ms":2407,"time_to_first_token_ms":351,"total_runtime_ms":2039,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0174,"confidence_interval":0.9393,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2343,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1828,"p95_latency_ms":2234,"p99_latency_ms":2396,"time_to_first_token_ms":327,"total_runtime_ms":2031,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0196,"confidence_interval":0.9428,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5474,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1848,"p95_latency_ms":2259,"p99_latency_ms":2423,"time_to_first_token_ms":336,"total_runtime_ms":2054,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0267,"confidence_interval":0.9465,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8438,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1844,"p95_latency_ms":2253,"p99_latency_ms":2417,"time_to_first_token_ms":325,"total_runtime_ms":2048,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0237,"confidence_interval":0.9323,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8915,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1849,"p95_latency_ms":2260,"p99_latency_ms":2425,"time_to_first_token_ms":349,"total_runtime_ms":2055,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0102,"confidence_interval":0.9387,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0302,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1831,"p95_latency_ms":2238,"p99_latency_ms":2401,"time_to_first_token_ms":322,"total_runtime_ms":2034,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.022,"confidence_interval":0.9496,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6031,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1100,"p95_latency_ms":1345,"p99_latency_ms":1443,"time_to_first_token_ms":297,"total_runtime_ms":1223,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0278,"confidence_interval":0.9394,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0519,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1031,"p95_latency_ms":1260,"p99_latency_ms":1351,"time_to_first_token_ms":298,"total_runtime_ms":1145,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0167,"confidence_interval":0.9305,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2439,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1062,"p95_latency_ms":1298,"p99_latency_ms":1393,"time_to_first_token_ms":283,"total_runtime_ms":1180,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0274,"confidence_interval":0.9474,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0225,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1082,"p95_latency_ms":1323,"p99_latency_ms":1419,"time_to_first_token_ms":275,"total_runtime_ms":1203,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0188,"confidence_interval":0.9373,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0267,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1079,"p95_latency_ms":1319,"p99_latency_ms":1414,"time_to_first_token_ms":300,"total_runtime_ms":1199,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0212,"confidence_interval":0.9304,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2615,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1060,"p95_latency_ms":1295,"p99_latency_ms":1389,"time_to_first_token_ms":281,"total_runtime_ms":1177,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0187,"confidence_interval":0.9381,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0564,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1079,"p95_latency_ms":1319,"p99_latency_ms":1415,"time_to_first_token_ms":299,"total_runtime_ms":1199,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0298,"confidence_interval":0.945,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9186,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1034,"p95_latency_ms":1264,"p99_latency_ms":1356,"time_to_first_token_ms":298,"total_runtime_ms":1149,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0276,"confidence_interval":0.9422,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3784,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1028,"p95_latency_ms":1256,"p99_latency_ms":1347,"time_to_first_token_ms":298,"total_runtime_ms":1142,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0233,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4345,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1118,"p95_latency_ms":1367,"p99_latency_ms":1466,"time_to_first_token_ms":299,"total_runtime_ms":1242,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0263,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0729,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2712,"p95_latency_ms":3315,"p99_latency_ms":3556,"time_to_first_token_ms":474,"total_runtime_ms":3014,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0147,"confidence_interval":0.9351,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2933,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2674,"p95_latency_ms":3268,"p99_latency_ms":3505,"time_to_first_token_ms":474,"total_runtime_ms":2971,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.017,"confidence_interval":0.9481,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.566,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2722,"p95_latency_ms":3327,"p99_latency_ms":3569,"time_to_first_token_ms":459,"total_runtime_ms":3024,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.012,"confidence_interval":0.9376,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3686,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2732,"p95_latency_ms":3339,"p99_latency_ms":3582,"time_to_first_token_ms":450,"total_runtime_ms":3035,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0135,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.69,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2716,"p95_latency_ms":3319,"p99_latency_ms":3560,"time_to_first_token_ms":472,"total_runtime_ms":3017,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0294,"confidence_interval":0.9459,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.917,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2684,"p95_latency_ms":3281,"p99_latency_ms":3519,"time_to_first_token_ms":474,"total_runtime_ms":2982,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.013,"confidence_interval":0.9306,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.934,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2739,"p95_latency_ms":3348,"p99_latency_ms":3591,"time_to_first_token_ms":457,"total_runtime_ms":3044,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0267,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0173,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2712,"p95_latency_ms":3314,"p99_latency_ms":3555,"time_to_first_token_ms":473,"total_runtime_ms":3013,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0162,"confidence_interval":0.946,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5785,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2707,"p95_latency_ms":3308,"p99_latency_ms":3549,"time_to_first_token_ms":473,"total_runtime_ms":3008,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0115,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1473,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2662,"p95_latency_ms":3254,"p99_latency_ms":3491,"time_to_first_token_ms":473,"total_runtime_ms":2958,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0145,"confidence_interval":0.9372,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8636,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1587,"p95_latency_ms":1940,"p99_latency_ms":2081,"time_to_first_token_ms":429,"total_runtime_ms":1764,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0271,"confidence_interval":0.9369,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0587,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1606,"p95_latency_ms":1962,"p99_latency_ms":2105,"time_to_first_token_ms":417,"total_runtime_ms":1784,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0148,"confidence_interval":0.9315,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.886,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1584,"p95_latency_ms":1936,"p99_latency_ms":2077,"time_to_first_token_ms":427,"total_runtime_ms":1760,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0104,"confidence_interval":0.9454,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.725,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1521,"p95_latency_ms":1859,"p99_latency_ms":1994,"time_to_first_token_ms":432,"total_runtime_ms":1690,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0221,"confidence_interval":0.9326,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2099,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1557,"p95_latency_ms":1903,"p99_latency_ms":2041,"time_to_first_token_ms":432,"total_runtime_ms":1730,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0242,"confidence_interval":0.9416,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1814,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1567,"p95_latency_ms":1915,"p99_latency_ms":2054,"time_to_first_token_ms":434,"total_runtime_ms":1741,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0223,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1428,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1540,"p95_latency_ms":1883,"p99_latency_ms":2020,"time_to_first_token_ms":404,"total_runtime_ms":1712,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0249,"confidence_interval":0.9358,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5671,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1612,"p95_latency_ms":1970,"p99_latency_ms":2113,"time_to_first_token_ms":404,"total_runtime_ms":1791,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.022,"confidence_interval":0.9331,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2084,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1608,"p95_latency_ms":1965,"p99_latency_ms":2108,"time_to_first_token_ms":431,"total_runtime_ms":1786,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0279,"confidence_interval":0.933,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2623,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1568,"p95_latency_ms":1916,"p99_latency_ms":2055,"time_to_first_token_ms":410,"total_runtime_ms":1742,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0115,"confidence_interval":0.9333,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1469,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":594,"p95_latency_ms":726,"p99_latency_ms":779,"time_to_first_token_ms":149,"total_runtime_ms":660,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.028,"confidence_interval":0.9466,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8992,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":555,"p95_latency_ms":678,"p99_latency_ms":727,"time_to_first_token_ms":134,"total_runtime_ms":616,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0123,"confidence_interval":0.9329,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4173,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":601,"p95_latency_ms":735,"p99_latency_ms":788,"time_to_first_token_ms":163,"total_runtime_ms":668,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0239,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8803,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":560,"p95_latency_ms":685,"p99_latency_ms":735,"time_to_first_token_ms":164,"total_runtime_ms":623,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0181,"confidence_interval":0.9357,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8966,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":529,"p95_latency_ms":646,"p99_latency_ms":693,"time_to_first_token_ms":147,"total_runtime_ms":588,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0282,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.967,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":505,"p95_latency_ms":617,"p99_latency_ms":662,"time_to_first_token_ms":156,"total_runtime_ms":561,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0267,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0189,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":600,"p95_latency_ms":733,"p99_latency_ms":786,"time_to_first_token_ms":160,"total_runtime_ms":666,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0108,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5028,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":596,"p95_latency_ms":728,"p99_latency_ms":781,"time_to_first_token_ms":140,"total_runtime_ms":662,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0176,"confidence_interval":0.9319,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1902,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":525,"p95_latency_ms":641,"p99_latency_ms":688,"time_to_first_token_ms":133,"total_runtime_ms":583,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0141,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5664,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":779,"time_to_first_token_ms":141,"total_runtime_ms":661,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0173,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.382,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":319,"p95_latency_ms":390,"p99_latency_ms":418,"time_to_first_token_ms":112,"total_runtime_ms":355,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0197,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.563,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":300,"p95_latency_ms":367,"p99_latency_ms":393,"time_to_first_token_ms":94,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9368,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1085,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":294,"p95_latency_ms":359,"p99_latency_ms":385,"time_to_first_token_ms":108,"total_runtime_ms":326,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0126,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5412,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":384,"p95_latency_ms":469,"p99_latency_ms":504,"time_to_first_token_ms":97,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0169,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2537,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":388,"p95_latency_ms":475,"p99_latency_ms":509,"time_to_first_token_ms":92,"total_runtime_ms":431,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0277,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4716,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":425,"time_to_first_token_ms":108,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0115,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1382,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":298,"p95_latency_ms":364,"p99_latency_ms":390,"time_to_first_token_ms":119,"total_runtime_ms":331,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0222,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9179,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":325,"p95_latency_ms":398,"p99_latency_ms":427,"time_to_first_token_ms":111,"total_runtime_ms":362,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0201,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7381,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":349,"p95_latency_ms":427,"p99_latency_ms":458,"time_to_first_token_ms":105,"total_runtime_ms":388,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0231,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6793,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":471,"time_to_first_token_ms":92,"total_runtime_ms":399,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0101,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.139,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":746,"p95_latency_ms":911,"p99_latency_ms":978,"time_to_first_token_ms":192,"total_runtime_ms":828,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9385,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3712,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":674,"p95_latency_ms":823,"p99_latency_ms":883,"time_to_first_token_ms":180,"total_runtime_ms":749,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9419,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2487,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":717,"p95_latency_ms":877,"p99_latency_ms":940,"time_to_first_token_ms":178,"total_runtime_ms":797,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0257,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9701,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":700,"p95_latency_ms":856,"p99_latency_ms":918,"time_to_first_token_ms":189,"total_runtime_ms":778,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0293,"confidence_interval":0.9483,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6574,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":673,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":171,"total_runtime_ms":747,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0153,"confidence_interval":0.9387,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6563,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":745,"p95_latency_ms":911,"p99_latency_ms":977,"time_to_first_token_ms":177,"total_runtime_ms":828,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0183,"confidence_interval":0.9313,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0372,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":750,"p95_latency_ms":917,"p99_latency_ms":984,"time_to_first_token_ms":187,"total_runtime_ms":834,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0223,"confidence_interval":0.9369,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5154,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":721,"p95_latency_ms":882,"p99_latency_ms":946,"time_to_first_token_ms":165,"total_runtime_ms":802,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0127,"confidence_interval":0.9355,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3993,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":706,"p95_latency_ms":863,"p99_latency_ms":926,"time_to_first_token_ms":186,"total_runtime_ms":784,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0258,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0574,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":740,"p95_latency_ms":904,"p99_latency_ms":970,"time_to_first_token_ms":164,"total_runtime_ms":822,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":49,"gpu_utilization_percent":36,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0247,"confidence_interval":0.9307,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6399,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":400,"p95_latency_ms":489,"p99_latency_ms":525,"time_to_first_token_ms":114,"total_runtime_ms":445,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0185,"confidence_interval":0.9461,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7068,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":446,"p95_latency_ms":545,"p99_latency_ms":584,"time_to_first_token_ms":113,"total_runtime_ms":495,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.011,"confidence_interval":0.9447,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1247,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":460,"p95_latency_ms":562,"p99_latency_ms":603,"time_to_first_token_ms":127,"total_runtime_ms":511,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0154,"confidence_interval":0.9474,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9989,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":416,"p95_latency_ms":508,"p99_latency_ms":545,"time_to_first_token_ms":143,"total_runtime_ms":462,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9325,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":480,"p95_latency_ms":587,"p99_latency_ms":630,"time_to_first_token_ms":119,"total_runtime_ms":534,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0242,"confidence_interval":0.9404,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.523,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":477,"p95_latency_ms":583,"p99_latency_ms":625,"time_to_first_token_ms":115,"total_runtime_ms":530,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.015,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5061,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":409,"p95_latency_ms":499,"p99_latency_ms":536,"time_to_first_token_ms":120,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0253,"confidence_interval":0.9432,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9614,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":577,"time_to_first_token_ms":130,"total_runtime_ms":489,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0106,"confidence_interval":0.9382,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5209,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":468,"p95_latency_ms":572,"p99_latency_ms":613,"time_to_first_token_ms":122,"total_runtime_ms":520,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0263,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0822,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":444,"p95_latency_ms":542,"p99_latency_ms":582,"time_to_first_token_ms":139,"total_runtime_ms":493,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":36,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.016,"confidence_interval":0.9364,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1987,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1345,"p95_latency_ms":1644,"p99_latency_ms":1764,"time_to_first_token_ms":265,"total_runtime_ms":1495,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9381,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4467,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1398,"p95_latency_ms":1709,"p99_latency_ms":1833,"time_to_first_token_ms":260,"total_runtime_ms":1554,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0214,"confidence_interval":0.9379,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8174,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1436,"p95_latency_ms":1755,"p99_latency_ms":1882,"time_to_first_token_ms":264,"total_runtime_ms":1595,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0228,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8192,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1369,"p95_latency_ms":1673,"p99_latency_ms":1794,"time_to_first_token_ms":285,"total_runtime_ms":1521,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0245,"confidence_interval":0.9485,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0385,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1344,"p95_latency_ms":1642,"p99_latency_ms":1762,"time_to_first_token_ms":267,"total_runtime_ms":1493,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0224,"confidence_interval":0.9377,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7803,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1359,"p95_latency_ms":1661,"p99_latency_ms":1781,"time_to_first_token_ms":258,"total_runtime_ms":1510,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0125,"confidence_interval":0.9359,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9247,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1435,"p95_latency_ms":1754,"p99_latency_ms":1882,"time_to_first_token_ms":277,"total_runtime_ms":1595,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0206,"confidence_interval":0.9311,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5631,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1373,"p95_latency_ms":1678,"p99_latency_ms":1800,"time_to_first_token_ms":273,"total_runtime_ms":1525,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.019,"confidence_interval":0.9376,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4309,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1399,"p95_latency_ms":1710,"p99_latency_ms":1834,"time_to_first_token_ms":285,"total_runtime_ms":1554,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0216,"confidence_interval":0.9331,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0211,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1440,"p95_latency_ms":1761,"p99_latency_ms":1889,"time_to_first_token_ms":272,"total_runtime_ms":1601,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":49,"gpu_utilization_percent":73,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0269,"confidence_interval":0.9432,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7297,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":822,"p95_latency_ms":1004,"p99_latency_ms":1077,"time_to_first_token_ms":231,"total_runtime_ms":913,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0236,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.754,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":873,"p95_latency_ms":1067,"p99_latency_ms":1145,"time_to_first_token_ms":220,"total_runtime_ms":970,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0203,"confidence_interval":0.9346,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.307,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":216,"total_runtime_ms":889,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0235,"confidence_interval":0.9494,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4531,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":769,"p95_latency_ms":940,"p99_latency_ms":1008,"time_to_first_token_ms":218,"total_runtime_ms":854,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0162,"confidence_interval":0.9321,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6824,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":782,"p95_latency_ms":956,"p99_latency_ms":1025,"time_to_first_token_ms":237,"total_runtime_ms":869,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0167,"confidence_interval":0.9362,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5179,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":791,"p95_latency_ms":967,"p99_latency_ms":1038,"time_to_first_token_ms":225,"total_runtime_ms":879,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0206,"confidence_interval":0.9385,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3866,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":871,"p95_latency_ms":1064,"p99_latency_ms":1141,"time_to_first_token_ms":212,"total_runtime_ms":967,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.011,"confidence_interval":0.9413,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1592,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":783,"p95_latency_ms":957,"p99_latency_ms":1026,"time_to_first_token_ms":236,"total_runtime_ms":870,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.022,"confidence_interval":0.9456,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.451,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":775,"p95_latency_ms":948,"p99_latency_ms":1017,"time_to_first_token_ms":226,"total_runtime_ms":861,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.021,"confidence_interval":0.9429,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6644,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":836,"p95_latency_ms":1022,"p99_latency_ms":1096,"time_to_first_token_ms":215,"total_runtime_ms":929,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":73,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0224,"confidence_interval":0.9372,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1297,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2039,"p95_latency_ms":2492,"p99_latency_ms":2673,"time_to_first_token_ms":372,"total_runtime_ms":2265,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0193,"confidence_interval":0.9415,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2731,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2004,"p95_latency_ms":2449,"p99_latency_ms":2627,"time_to_first_token_ms":350,"total_runtime_ms":2226,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0173,"confidence_interval":0.9341,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.314,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1995,"p95_latency_ms":2438,"p99_latency_ms":2616,"time_to_first_token_ms":356,"total_runtime_ms":2217,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0144,"confidence_interval":0.9361,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4994,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2053,"p95_latency_ms":2509,"p99_latency_ms":2692,"time_to_first_token_ms":368,"total_runtime_ms":2281,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0185,"confidence_interval":0.9306,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5826,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2065,"p95_latency_ms":2524,"p99_latency_ms":2708,"time_to_first_token_ms":358,"total_runtime_ms":2295,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.9374,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3427,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2042,"p95_latency_ms":2495,"p99_latency_ms":2677,"time_to_first_token_ms":355,"total_runtime_ms":2269,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0125,"confidence_interval":0.9402,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4045,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2007,"p95_latency_ms":2453,"p99_latency_ms":2631,"time_to_first_token_ms":351,"total_runtime_ms":2230,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0136,"confidence_interval":0.9393,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0479,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2056,"p95_latency_ms":2513,"p99_latency_ms":2696,"time_to_first_token_ms":355,"total_runtime_ms":2285,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0272,"confidence_interval":0.9408,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7081,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2071,"p95_latency_ms":2532,"p99_latency_ms":2716,"time_to_first_token_ms":359,"total_runtime_ms":2301,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0104,"confidence_interval":0.9313,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3833,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2032,"p95_latency_ms":2484,"p99_latency_ms":2664,"time_to_first_token_ms":380,"total_runtime_ms":2258,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":49,"gpu_utilization_percent":100,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0152,"confidence_interval":0.9385,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1679,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1167,"p95_latency_ms":1426,"p99_latency_ms":1530,"time_to_first_token_ms":321,"total_runtime_ms":1297,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0296,"confidence_interval":0.9439,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7076,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1169,"p95_latency_ms":1429,"p99_latency_ms":1533,"time_to_first_token_ms":337,"total_runtime_ms":1299,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0163,"confidence_interval":0.9361,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6835,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1183,"p95_latency_ms":1446,"p99_latency_ms":1551,"time_to_first_token_ms":313,"total_runtime_ms":1315,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.019,"confidence_interval":0.9409,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1211,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1220,"p95_latency_ms":1491,"p99_latency_ms":1600,"time_to_first_token_ms":315,"total_runtime_ms":1356,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0165,"confidence_interval":0.9458,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2024,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1194,"p95_latency_ms":1459,"p99_latency_ms":1566,"time_to_first_token_ms":337,"total_runtime_ms":1327,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0182,"confidence_interval":0.9329,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1502,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1156,"p95_latency_ms":1413,"p99_latency_ms":1515,"time_to_first_token_ms":329,"total_runtime_ms":1284,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0196,"confidence_interval":0.9447,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7382,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1158,"p95_latency_ms":1416,"p99_latency_ms":1518,"time_to_first_token_ms":333,"total_runtime_ms":1287,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0157,"confidence_interval":0.94,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3799,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1177,"p95_latency_ms":1438,"p99_latency_ms":1543,"time_to_first_token_ms":317,"total_runtime_ms":1308,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0274,"confidence_interval":0.9343,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7528,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1190,"p95_latency_ms":1454,"p99_latency_ms":1560,"time_to_first_token_ms":333,"total_runtime_ms":1322,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0235,"confidence_interval":0.9404,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8621,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1147,"p95_latency_ms":1402,"p99_latency_ms":1504,"time_to_first_token_ms":323,"total_runtime_ms":1274,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":100,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.022,"confidence_interval":0.9414,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3769,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":672,"time_to_first_token_ms":135,"total_runtime_ms":569,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0206,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5607,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":456,"p95_latency_ms":557,"p99_latency_ms":598,"time_to_first_token_ms":152,"total_runtime_ms":506,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0224,"confidence_interval":0.9434,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3441,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":473,"p95_latency_ms":578,"p99_latency_ms":620,"time_to_first_token_ms":138,"total_runtime_ms":526,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0204,"confidence_interval":0.9475,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6768,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":485,"p95_latency_ms":593,"p99_latency_ms":636,"time_to_first_token_ms":134,"total_runtime_ms":539,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0194,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2936,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":448,"p95_latency_ms":548,"p99_latency_ms":588,"time_to_first_token_ms":142,"total_runtime_ms":498,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0143,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0434,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":476,"p95_latency_ms":582,"p99_latency_ms":624,"time_to_first_token_ms":125,"total_runtime_ms":529,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0138,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3402,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":490,"p95_latency_ms":599,"p99_latency_ms":643,"time_to_first_token_ms":130,"total_runtime_ms":545,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0258,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1191,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":485,"p95_latency_ms":593,"p99_latency_ms":636,"time_to_first_token_ms":153,"total_runtime_ms":539,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0186,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8501,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":429,"p95_latency_ms":525,"p99_latency_ms":563,"time_to_first_token_ms":137,"total_runtime_ms":477,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0164,"confidence_interval":0.9348,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2289,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":427,"p95_latency_ms":522,"p99_latency_ms":560,"time_to_first_token_ms":138,"total_runtime_ms":475,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0286,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1746,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":325,"p95_latency_ms":397,"p99_latency_ms":426,"time_to_first_token_ms":84,"total_runtime_ms":361,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1364,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":293,"p95_latency_ms":358,"p99_latency_ms":384,"time_to_first_token_ms":92,"total_runtime_ms":325,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0151,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4007,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":357,"p99_latency_ms":383,"time_to_first_token_ms":88,"total_runtime_ms":325,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0147,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8337,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":344,"p95_latency_ms":421,"p99_latency_ms":451,"time_to_first_token_ms":86,"total_runtime_ms":382,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0117,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4076,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":339,"p95_latency_ms":414,"p99_latency_ms":445,"time_to_first_token_ms":83,"total_runtime_ms":377,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0133,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3714,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":262,"p95_latency_ms":320,"p99_latency_ms":343,"time_to_first_token_ms":94,"total_runtime_ms":291,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.018,"confidence_interval":0.9448,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3287,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":338,"p95_latency_ms":413,"p99_latency_ms":443,"time_to_first_token_ms":103,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.011,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9321,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":424,"time_to_first_token_ms":77,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0175,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9305,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":263,"p95_latency_ms":321,"p99_latency_ms":345,"time_to_first_token_ms":80,"total_runtime_ms":292,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0293,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4335,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":408,"time_to_first_token_ms":91,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0248,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2064,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":565,"p95_latency_ms":690,"p99_latency_ms":741,"time_to_first_token_ms":171,"total_runtime_ms":628,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0255,"confidence_interval":0.9402,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9907,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":572,"p95_latency_ms":700,"p99_latency_ms":750,"time_to_first_token_ms":156,"total_runtime_ms":636,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0165,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1533,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":573,"p95_latency_ms":701,"p99_latency_ms":751,"time_to_first_token_ms":151,"total_runtime_ms":637,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0158,"confidence_interval":0.9469,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7745,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":604,"p95_latency_ms":738,"p99_latency_ms":792,"time_to_first_token_ms":171,"total_runtime_ms":671,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0152,"confidence_interval":0.9412,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3622,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":569,"p95_latency_ms":696,"p99_latency_ms":746,"time_to_first_token_ms":167,"total_runtime_ms":633,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0233,"confidence_interval":0.9372,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7245,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":648,"p95_latency_ms":792,"p99_latency_ms":850,"time_to_first_token_ms":174,"total_runtime_ms":720,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0169,"confidence_interval":0.9375,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4995,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":621,"p95_latency_ms":759,"p99_latency_ms":814,"time_to_first_token_ms":171,"total_runtime_ms":690,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0139,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8953,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":654,"p95_latency_ms":799,"p99_latency_ms":857,"time_to_first_token_ms":150,"total_runtime_ms":726,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0141,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2285,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":561,"p95_latency_ms":686,"p99_latency_ms":736,"time_to_first_token_ms":143,"total_runtime_ms":623,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.011,"confidence_interval":0.9477,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6725,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":614,"p95_latency_ms":751,"p99_latency_ms":805,"time_to_first_token_ms":156,"total_runtime_ms":683,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0261,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1019,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":335,"p95_latency_ms":409,"p99_latency_ms":439,"time_to_first_token_ms":103,"total_runtime_ms":372,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.9371,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3427,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":381,"p95_latency_ms":466,"p99_latency_ms":500,"time_to_first_token_ms":107,"total_runtime_ms":424,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.9389,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1483,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":478,"time_to_first_token_ms":112,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0254,"confidence_interval":0.932,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.921,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":376,"p95_latency_ms":460,"p99_latency_ms":493,"time_to_first_token_ms":108,"total_runtime_ms":418,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0202,"confidence_interval":0.9324,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2769,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":361,"p95_latency_ms":442,"p99_latency_ms":474,"time_to_first_token_ms":103,"total_runtime_ms":401,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0251,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9239,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":496,"time_to_first_token_ms":127,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0199,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9202,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":352,"p95_latency_ms":430,"p99_latency_ms":461,"time_to_first_token_ms":124,"total_runtime_ms":391,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0168,"confidence_interval":0.9389,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2237,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":361,"p95_latency_ms":442,"p99_latency_ms":474,"time_to_first_token_ms":97,"total_runtime_ms":402,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0197,"confidence_interval":0.9334,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0571,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":381,"p95_latency_ms":466,"p99_latency_ms":500,"time_to_first_token_ms":107,"total_runtime_ms":423,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0122,"confidence_interval":0.9402,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5849,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":414,"p95_latency_ms":506,"p99_latency_ms":543,"time_to_first_token_ms":129,"total_runtime_ms":460,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0114,"confidence_interval":0.9335,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3009,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1138,"p95_latency_ms":1390,"p99_latency_ms":1491,"time_to_first_token_ms":228,"total_runtime_ms":1264,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0171,"confidence_interval":0.9306,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6533,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1142,"p95_latency_ms":1395,"p99_latency_ms":1497,"time_to_first_token_ms":250,"total_runtime_ms":1268,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.028,"confidence_interval":0.9401,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0478,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1159,"p95_latency_ms":1416,"p99_latency_ms":1519,"time_to_first_token_ms":255,"total_runtime_ms":1288,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0252,"confidence_interval":0.946,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8704,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1150,"p95_latency_ms":1406,"p99_latency_ms":1508,"time_to_first_token_ms":243,"total_runtime_ms":1278,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0116,"confidence_interval":0.9469,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5411,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1149,"p95_latency_ms":1404,"p99_latency_ms":1506,"time_to_first_token_ms":223,"total_runtime_ms":1276,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0216,"confidence_interval":0.9489,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8623,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1165,"p95_latency_ms":1423,"p99_latency_ms":1527,"time_to_first_token_ms":242,"total_runtime_ms":1294,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0185,"confidence_interval":0.9333,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.44,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1126,"p95_latency_ms":1376,"p99_latency_ms":1477,"time_to_first_token_ms":235,"total_runtime_ms":1251,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0166,"confidence_interval":0.9367,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2969,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1221,"p95_latency_ms":1492,"p99_latency_ms":1600,"time_to_first_token_ms":242,"total_runtime_ms":1356,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0225,"confidence_interval":0.9352,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9809,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1218,"p95_latency_ms":1488,"p99_latency_ms":1597,"time_to_first_token_ms":235,"total_runtime_ms":1353,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.017,"confidence_interval":0.9329,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3668,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1214,"p95_latency_ms":1484,"p99_latency_ms":1592,"time_to_first_token_ms":254,"total_runtime_ms":1349,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0268,"confidence_interval":0.9408,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8213,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":686,"p95_latency_ms":838,"p99_latency_ms":899,"time_to_first_token_ms":178,"total_runtime_ms":762,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0253,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1242,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":715,"p95_latency_ms":874,"p99_latency_ms":938,"time_to_first_token_ms":203,"total_runtime_ms":795,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0218,"confidence_interval":0.9369,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6115,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":714,"p95_latency_ms":872,"p99_latency_ms":935,"time_to_first_token_ms":211,"total_runtime_ms":793,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0245,"confidence_interval":0.9399,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0056,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":742,"p95_latency_ms":907,"p99_latency_ms":973,"time_to_first_token_ms":212,"total_runtime_ms":825,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.019,"confidence_interval":0.9413,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5275,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":666,"p95_latency_ms":814,"p99_latency_ms":873,"time_to_first_token_ms":177,"total_runtime_ms":740,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0111,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4718,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":727,"p95_latency_ms":888,"p99_latency_ms":953,"time_to_first_token_ms":186,"total_runtime_ms":808,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0193,"confidence_interval":0.9351,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4701,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":704,"p95_latency_ms":860,"p99_latency_ms":923,"time_to_first_token_ms":200,"total_runtime_ms":782,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0131,"confidence_interval":0.9484,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1899,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":645,"p95_latency_ms":788,"p99_latency_ms":846,"time_to_first_token_ms":188,"total_runtime_ms":717,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0141,"confidence_interval":0.9473,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2933,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":747,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":198,"total_runtime_ms":831,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.016,"confidence_interval":0.9492,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0993,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:7b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":642,"p95_latency_ms":785,"p99_latency_ms":842,"time_to_first_token_ms":206,"total_runtime_ms":713,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0207,"confidence_interval":0.9308,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.387,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1706,"p95_latency_ms":2085,"p99_latency_ms":2236,"time_to_first_token_ms":330,"total_runtime_ms":1895,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0152,"confidence_interval":0.938,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.644,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1751,"p95_latency_ms":2140,"p99_latency_ms":2296,"time_to_first_token_ms":313,"total_runtime_ms":1946,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0204,"confidence_interval":0.9335,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9212,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1698,"p95_latency_ms":2076,"p99_latency_ms":2227,"time_to_first_token_ms":310,"total_runtime_ms":1887,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0138,"confidence_interval":0.9484,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2539,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1736,"p95_latency_ms":2122,"p99_latency_ms":2276,"time_to_first_token_ms":309,"total_runtime_ms":1929,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0142,"confidence_interval":0.9409,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9426,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1672,"p95_latency_ms":2043,"p99_latency_ms":2192,"time_to_first_token_ms":321,"total_runtime_ms":1858,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0168,"confidence_interval":0.9488,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2188,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1689,"p95_latency_ms":2064,"p99_latency_ms":2214,"time_to_first_token_ms":307,"total_runtime_ms":1877,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0195,"confidence_interval":0.9368,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2673,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1666,"p95_latency_ms":2037,"p99_latency_ms":2185,"time_to_first_token_ms":316,"total_runtime_ms":1852,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0173,"confidence_interval":0.9396,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6338,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1705,"p95_latency_ms":2084,"p99_latency_ms":2236,"time_to_first_token_ms":311,"total_runtime_ms":1895,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0213,"confidence_interval":0.9381,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6169,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1694,"p95_latency_ms":2070,"p99_latency_ms":2221,"time_to_first_token_ms":313,"total_runtime_ms":1882,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0268,"confidence_interval":0.9495,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5569,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1657,"p95_latency_ms":2025,"p99_latency_ms":2172,"time_to_first_token_ms":313,"total_runtime_ms":1841,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0161,"confidence_interval":0.9458,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9468,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1040,"p95_latency_ms":1272,"p99_latency_ms":1364,"time_to_first_token_ms":284,"total_runtime_ms":1156,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9415,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7157,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1014,"p95_latency_ms":1240,"p99_latency_ms":1330,"time_to_first_token_ms":273,"total_runtime_ms":1127,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0172,"confidence_interval":0.933,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7228,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1030,"p95_latency_ms":1258,"p99_latency_ms":1350,"time_to_first_token_ms":266,"total_runtime_ms":1144,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0206,"confidence_interval":0.9318,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7683,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1033,"p95_latency_ms":1262,"p99_latency_ms":1354,"time_to_first_token_ms":276,"total_runtime_ms":1147,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0232,"confidence_interval":0.9491,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.752,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":994,"p95_latency_ms":1215,"p99_latency_ms":1303,"time_to_first_token_ms":283,"total_runtime_ms":1104,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0274,"confidence_interval":0.9474,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8098,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1023,"p95_latency_ms":1251,"p99_latency_ms":1342,"time_to_first_token_ms":288,"total_runtime_ms":1137,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0113,"confidence_interval":0.9389,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4192,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":953,"p95_latency_ms":1165,"p99_latency_ms":1250,"time_to_first_token_ms":287,"total_runtime_ms":1059,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.026,"confidence_interval":0.95,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3091,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1030,"p95_latency_ms":1259,"p99_latency_ms":1350,"time_to_first_token_ms":281,"total_runtime_ms":1144,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0227,"confidence_interval":0.9491,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4769,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":986,"p95_latency_ms":1205,"p99_latency_ms":1292,"time_to_first_token_ms":265,"total_runtime_ms":1095,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0147,"confidence_interval":0.9473,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7388,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"contradictory_intent|high|qwen2:14b|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":956,"p95_latency_ms":1169,"p99_latency_ms":1254,"time_to_first_token_ms":288,"total_runtime_ms":1062,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.025,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5701,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1130,"p95_latency_ms":1381,"p99_latency_ms":1481,"time_to_first_token_ms":210,"total_runtime_ms":1255,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0294,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4829,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1111,"p95_latency_ms":1358,"p99_latency_ms":1457,"time_to_first_token_ms":220,"total_runtime_ms":1235,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.025,"confidence_interval":0.9495,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3699,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1096,"p95_latency_ms":1340,"p99_latency_ms":1437,"time_to_first_token_ms":193,"total_runtime_ms":1218,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0212,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4239,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1074,"p95_latency_ms":1312,"p99_latency_ms":1408,"time_to_first_token_ms":196,"total_runtime_ms":1193,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0187,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3114,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1079,"p95_latency_ms":1319,"p99_latency_ms":1415,"time_to_first_token_ms":197,"total_runtime_ms":1199,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0243,"confidence_interval":0.9481,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4141,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1134,"p95_latency_ms":1386,"p99_latency_ms":1487,"time_to_first_token_ms":202,"total_runtime_ms":1260,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0203,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3784,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1147,"p95_latency_ms":1402,"p99_latency_ms":1504,"time_to_first_token_ms":188,"total_runtime_ms":1274,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0145,"confidence_interval":0.9466,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8545,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1095,"p95_latency_ms":1338,"p99_latency_ms":1435,"time_to_first_token_ms":213,"total_runtime_ms":1216,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0169,"confidence_interval":0.9445,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4206,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1049,"p95_latency_ms":1282,"p99_latency_ms":1375,"time_to_first_token_ms":221,"total_runtime_ms":1165,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0117,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4655,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1063,"p95_latency_ms":1300,"p99_latency_ms":1394,"time_to_first_token_ms":194,"total_runtime_ms":1182,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0104,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8153,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":778,"p95_latency_ms":951,"p99_latency_ms":1020,"time_to_first_token_ms":151,"total_runtime_ms":864,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0289,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5922,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":716,"p95_latency_ms":876,"p99_latency_ms":939,"time_to_first_token_ms":144,"total_runtime_ms":796,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0101,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2755,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":703,"p95_latency_ms":859,"p99_latency_ms":922,"time_to_first_token_ms":173,"total_runtime_ms":781,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0199,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2788,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":727,"p95_latency_ms":889,"p99_latency_ms":953,"time_to_first_token_ms":158,"total_runtime_ms":808,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0232,"confidence_interval":0.9416,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4631,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":694,"p95_latency_ms":849,"p99_latency_ms":910,"time_to_first_token_ms":149,"total_runtime_ms":771,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0269,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1276,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":729,"p95_latency_ms":891,"p99_latency_ms":956,"time_to_first_token_ms":161,"total_runtime_ms":810,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0264,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3392,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":715,"p95_latency_ms":873,"p99_latency_ms":937,"time_to_first_token_ms":149,"total_runtime_ms":794,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0165,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1386,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":694,"p95_latency_ms":848,"p99_latency_ms":909,"time_to_first_token_ms":161,"total_runtime_ms":771,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0125,"confidence_interval":0.9321,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0056,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":767,"p95_latency_ms":937,"p99_latency_ms":1005,"time_to_first_token_ms":176,"total_runtime_ms":852,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0197,"confidence_interval":0.9438,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4307,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":709,"p95_latency_ms":866,"p99_latency_ms":929,"time_to_first_token_ms":165,"total_runtime_ms":788,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0202,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8052,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1496,"p95_latency_ms":1828,"p99_latency_ms":1961,"time_to_first_token_ms":250,"total_runtime_ms":1662,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0168,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2892,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1512,"p95_latency_ms":1848,"p99_latency_ms":1983,"time_to_first_token_ms":257,"total_runtime_ms":1680,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9334,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6542,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1489,"p95_latency_ms":1820,"p99_latency_ms":1952,"time_to_first_token_ms":241,"total_runtime_ms":1654,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0254,"confidence_interval":0.9423,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2197,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1418,"p95_latency_ms":1733,"p99_latency_ms":1859,"time_to_first_token_ms":242,"total_runtime_ms":1576,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0167,"confidence_interval":0.9497,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7325,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1472,"p95_latency_ms":1799,"p99_latency_ms":1929,"time_to_first_token_ms":262,"total_runtime_ms":1635,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0219,"confidence_interval":0.942,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9977,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1460,"p95_latency_ms":1784,"p99_latency_ms":1914,"time_to_first_token_ms":256,"total_runtime_ms":1622,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0214,"confidence_interval":0.9463,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0575,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1486,"p95_latency_ms":1816,"p99_latency_ms":1948,"time_to_first_token_ms":239,"total_runtime_ms":1651,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0156,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0014,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1454,"p95_latency_ms":1778,"p99_latency_ms":1907,"time_to_first_token_ms":252,"total_runtime_ms":1616,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0113,"confidence_interval":0.936,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9351,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1506,"p95_latency_ms":1841,"p99_latency_ms":1975,"time_to_first_token_ms":233,"total_runtime_ms":1674,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0461,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1496,"p95_latency_ms":1828,"p99_latency_ms":1961,"time_to_first_token_ms":260,"total_runtime_ms":1662,"cold_start_penalty_ms":562,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.024,"confidence_interval":0.934,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.6491,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":984,"p95_latency_ms":1203,"p99_latency_ms":1290,"time_to_first_token_ms":188,"total_runtime_ms":1093,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0112,"confidence_interval":0.936,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.054,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":978,"p95_latency_ms":1196,"p99_latency_ms":1282,"time_to_first_token_ms":202,"total_runtime_ms":1087,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0274,"confidence_interval":0.9426,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0803,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":978,"p95_latency_ms":1195,"p99_latency_ms":1282,"time_to_first_token_ms":191,"total_runtime_ms":1086,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0221,"confidence_interval":0.9304,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5305,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":922,"p95_latency_ms":1126,"p99_latency_ms":1208,"time_to_first_token_ms":202,"total_runtime_ms":1024,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0292,"confidence_interval":0.9417,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3074,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":956,"p95_latency_ms":1168,"p99_latency_ms":1253,"time_to_first_token_ms":210,"total_runtime_ms":1062,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.013,"confidence_interval":0.9487,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4144,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":949,"p95_latency_ms":1160,"p99_latency_ms":1244,"time_to_first_token_ms":222,"total_runtime_ms":1054,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0266,"confidence_interval":0.9381,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7781,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":976,"p95_latency_ms":1193,"p99_latency_ms":1280,"time_to_first_token_ms":190,"total_runtime_ms":1085,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0122,"confidence_interval":0.9324,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7938,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":985,"p95_latency_ms":1204,"p99_latency_ms":1291,"time_to_first_token_ms":188,"total_runtime_ms":1094,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0154,"confidence_interval":0.9475,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3159,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1016,"p95_latency_ms":1242,"p99_latency_ms":1333,"time_to_first_token_ms":216,"total_runtime_ms":1129,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0172,"confidence_interval":0.9416,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4846,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":914,"p95_latency_ms":1117,"p99_latency_ms":1198,"time_to_first_token_ms":197,"total_runtime_ms":1016,"cold_start_penalty_ms":0,"peak_ram_mb":1633,"avg_ram_mb":1089,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.02,"confidence_interval":0.9334,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9406,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2900,"p95_latency_ms":3544,"p99_latency_ms":3802,"time_to_first_token_ms":436,"total_runtime_ms":3222,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0248,"confidence_interval":0.9384,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9997,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2888,"p95_latency_ms":3529,"p99_latency_ms":3786,"time_to_first_token_ms":412,"total_runtime_ms":3209,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.03,"confidence_interval":0.9472,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3498,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2886,"p95_latency_ms":3528,"p99_latency_ms":3784,"time_to_first_token_ms":436,"total_runtime_ms":3207,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.027,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5204,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2849,"p95_latency_ms":3482,"p99_latency_ms":3735,"time_to_first_token_ms":417,"total_runtime_ms":3165,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.011,"confidence_interval":0.9356,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2524,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2909,"p95_latency_ms":3555,"p99_latency_ms":3814,"time_to_first_token_ms":430,"total_runtime_ms":3232,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0156,"confidence_interval":0.9455,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0854,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2917,"p95_latency_ms":3565,"p99_latency_ms":3825,"time_to_first_token_ms":435,"total_runtime_ms":3241,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0148,"confidence_interval":0.9307,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5769,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2859,"p95_latency_ms":3494,"p99_latency_ms":3748,"time_to_first_token_ms":407,"total_runtime_ms":3177,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0185,"confidence_interval":0.9453,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0801,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2881,"p95_latency_ms":3521,"p99_latency_ms":3777,"time_to_first_token_ms":411,"total_runtime_ms":3201,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0231,"confidence_interval":0.9328,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1249,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2861,"p95_latency_ms":3497,"p99_latency_ms":3751,"time_to_first_token_ms":429,"total_runtime_ms":3179,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0175,"confidence_interval":0.9353,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.263,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2840,"p95_latency_ms":3471,"p99_latency_ms":3723,"time_to_first_token_ms":410,"total_runtime_ms":3155,"cold_start_penalty_ms":1123,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0262,"confidence_interval":0.9311,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4897,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1907,"p95_latency_ms":2331,"p99_latency_ms":2501,"time_to_first_token_ms":380,"total_runtime_ms":2119,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0182,"confidence_interval":0.9488,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1314,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1852,"p95_latency_ms":2263,"p99_latency_ms":2428,"time_to_first_token_ms":394,"total_runtime_ms":2057,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0206,"confidence_interval":0.9451,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2831,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1895,"p95_latency_ms":2317,"p99_latency_ms":2485,"time_to_first_token_ms":364,"total_runtime_ms":2106,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0268,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.281,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1911,"p95_latency_ms":2336,"p99_latency_ms":2506,"time_to_first_token_ms":395,"total_runtime_ms":2124,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.013,"confidence_interval":0.9429,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1222,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1865,"p95_latency_ms":2280,"p99_latency_ms":2445,"time_to_first_token_ms":381,"total_runtime_ms":2072,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.026,"confidence_interval":0.9429,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2318,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1823,"p95_latency_ms":2228,"p99_latency_ms":2390,"time_to_first_token_ms":388,"total_runtime_ms":2026,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0153,"confidence_interval":0.9452,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4891,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1865,"p95_latency_ms":2279,"p99_latency_ms":2445,"time_to_first_token_ms":393,"total_runtime_ms":2072,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.011,"confidence_interval":0.9477,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0275,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1857,"p95_latency_ms":2269,"p99_latency_ms":2435,"time_to_first_token_ms":381,"total_runtime_ms":2063,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9486,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6508,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1884,"p95_latency_ms":2303,"p99_latency_ms":2470,"time_to_first_token_ms":367,"total_runtime_ms":2093,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9483,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6502,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:7b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1882,"p95_latency_ms":2300,"p99_latency_ms":2467,"time_to_first_token_ms":369,"total_runtime_ms":2091,"cold_start_penalty_ms":0,"peak_ram_mb":3266,"avg_ram_mb":2177,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0177,"confidence_interval":0.9415,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.2681,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4290,"p95_latency_ms":5243,"p99_latency_ms":5625,"time_to_first_token_ms":594,"total_runtime_ms":4767,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0292,"confidence_interval":0.9479,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2297,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4207,"p95_latency_ms":5141,"p99_latency_ms":5515,"time_to_first_token_ms":579,"total_runtime_ms":4674,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.9411,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2265,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4209,"p95_latency_ms":5145,"p99_latency_ms":5519,"time_to_first_token_ms":582,"total_runtime_ms":4677,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0205,"confidence_interval":0.9316,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8136,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4298,"p95_latency_ms":5253,"p99_latency_ms":5635,"time_to_first_token_ms":578,"total_runtime_ms":4776,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0278,"confidence_interval":0.9387,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.1496,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4280,"p95_latency_ms":5231,"p99_latency_ms":5612,"time_to_first_token_ms":574,"total_runtime_ms":4756,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.026,"confidence_interval":0.9389,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6528,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4220,"p95_latency_ms":5158,"p99_latency_ms":5533,"time_to_first_token_ms":606,"total_runtime_ms":4689,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0181,"confidence_interval":0.9401,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8643,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4276,"p95_latency_ms":5226,"p99_latency_ms":5607,"time_to_first_token_ms":595,"total_runtime_ms":4751,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0263,"confidence_interval":0.9352,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5652,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4242,"p95_latency_ms":5185,"p99_latency_ms":5562,"time_to_first_token_ms":599,"total_runtime_ms":4714,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0179,"confidence_interval":0.9322,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3787,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4241,"p95_latency_ms":5183,"p99_latency_ms":5560,"time_to_first_token_ms":591,"total_runtime_ms":4712,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.023,"confidence_interval":0.9427,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.499,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":4278,"p95_latency_ms":5229,"p99_latency_ms":5610,"time_to_first_token_ms":576,"total_runtime_ms":4754,"cold_start_penalty_ms":1664,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0143,"confidence_interval":0.9401,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.6725,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2747,"p95_latency_ms":3357,"p99_latency_ms":3602,"time_to_first_token_ms":540,"total_runtime_ms":3052,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0261,"confidence_interval":0.9303,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5864,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2783,"p95_latency_ms":3401,"p99_latency_ms":3649,"time_to_first_token_ms":554,"total_runtime_ms":3092,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0182,"confidence_interval":0.9448,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0744,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2765,"p95_latency_ms":3379,"p99_latency_ms":3625,"time_to_first_token_ms":546,"total_runtime_ms":3072,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0267,"confidence_interval":0.9317,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.511,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2784,"p95_latency_ms":3403,"p99_latency_ms":3650,"time_to_first_token_ms":538,"total_runtime_ms":3093,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0165,"confidence_interval":0.9337,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9148,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2724,"p95_latency_ms":3329,"p99_latency_ms":3571,"time_to_first_token_ms":535,"total_runtime_ms":3027,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0289,"confidence_interval":0.9332,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3797,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2770,"p95_latency_ms":3385,"p99_latency_ms":3631,"time_to_first_token_ms":560,"total_runtime_ms":3077,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0227,"confidence_interval":0.9399,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2914,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2741,"p95_latency_ms":3350,"p99_latency_ms":3594,"time_to_first_token_ms":547,"total_runtime_ms":3046,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0192,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.3112,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2767,"p95_latency_ms":3381,"p99_latency_ms":3627,"time_to_first_token_ms":561,"total_runtime_ms":3074,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0245,"confidence_interval":0.9355,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2884,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2790,"p95_latency_ms":3410,"p99_latency_ms":3658,"time_to_first_token_ms":544,"total_runtime_ms":3100,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0285,"confidence_interval":0.9429,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8597,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|low|qwen2:14b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2796,"p95_latency_ms":3417,"p99_latency_ms":3666,"time_to_first_token_ms":527,"total_runtime_ms":3106,"cold_start_penalty_ms":0,"peak_ram_mb":4838,"avg_ram_mb":3226,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0268,"confidence_interval":0.9303,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7884,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":885,"p95_latency_ms":1082,"p99_latency_ms":1161,"time_to_first_token_ms":160,"total_runtime_ms":984,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0261,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1375,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":819,"p95_latency_ms":1001,"p99_latency_ms":1074,"time_to_first_token_ms":171,"total_runtime_ms":910,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1356,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":855,"p95_latency_ms":1045,"p99_latency_ms":1121,"time_to_first_token_ms":159,"total_runtime_ms":950,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0197,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9777,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":807,"p95_latency_ms":986,"p99_latency_ms":1058,"time_to_first_token_ms":185,"total_runtime_ms":897,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0123,"confidence_interval":0.9397,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4354,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1052,"time_to_first_token_ms":190,"total_runtime_ms":892,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0295,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8259,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":810,"p95_latency_ms":990,"p99_latency_ms":1062,"time_to_first_token_ms":171,"total_runtime_ms":900,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0297,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0482,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":828,"p95_latency_ms":1012,"p99_latency_ms":1086,"time_to_first_token_ms":189,"total_runtime_ms":920,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0159,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9145,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":788,"p95_latency_ms":964,"p99_latency_ms":1034,"time_to_first_token_ms":182,"total_runtime_ms":876,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.028,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9192,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":856,"p95_latency_ms":1046,"p99_latency_ms":1122,"time_to_first_token_ms":189,"total_runtime_ms":951,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.027,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1552,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":799,"p95_latency_ms":977,"p99_latency_ms":1048,"time_to_first_token_ms":171,"total_runtime_ms":888,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0167,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8524,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":786,"time_to_first_token_ms":132,"total_runtime_ms":666,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0225,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6235,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":610,"p95_latency_ms":745,"p99_latency_ms":800,"time_to_first_token_ms":119,"total_runtime_ms":678,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.015,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6694,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":587,"p95_latency_ms":717,"p99_latency_ms":769,"time_to_first_token_ms":144,"total_runtime_ms":652,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0113,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9955,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":559,"p95_latency_ms":683,"p99_latency_ms":732,"time_to_first_token_ms":139,"total_runtime_ms":621,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0137,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3206,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":552,"p95_latency_ms":675,"p99_latency_ms":724,"time_to_first_token_ms":121,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0287,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6525,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":756,"time_to_first_token_ms":125,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0108,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.91,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":542,"p95_latency_ms":662,"p99_latency_ms":711,"time_to_first_token_ms":122,"total_runtime_ms":602,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0265,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3112,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":547,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":145,"total_runtime_ms":608,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.013,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0058,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":568,"p95_latency_ms":695,"p99_latency_ms":745,"time_to_first_token_ms":124,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0158,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2569,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":542,"p95_latency_ms":662,"p99_latency_ms":710,"time_to_first_token_ms":128,"total_runtime_ms":602,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0291,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6516,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1092,"p95_latency_ms":1334,"p99_latency_ms":1431,"time_to_first_token_ms":215,"total_runtime_ms":1213,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.9413,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.017,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1131,"p95_latency_ms":1382,"p99_latency_ms":1482,"time_to_first_token_ms":194,"total_runtime_ms":1256,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0157,"confidence_interval":0.9492,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5546,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1169,"p95_latency_ms":1428,"p99_latency_ms":1532,"time_to_first_token_ms":216,"total_runtime_ms":1299,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0249,"confidence_interval":0.9431,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.995,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1066,"p95_latency_ms":1303,"p99_latency_ms":1398,"time_to_first_token_ms":214,"total_runtime_ms":1184,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9395,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.265,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1119,"p95_latency_ms":1368,"p99_latency_ms":1467,"time_to_first_token_ms":212,"total_runtime_ms":1243,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0114,"confidence_interval":0.9359,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1243,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1138,"p95_latency_ms":1391,"p99_latency_ms":1492,"time_to_first_token_ms":210,"total_runtime_ms":1265,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.026,"confidence_interval":0.941,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2152,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1063,"p95_latency_ms":1300,"p99_latency_ms":1394,"time_to_first_token_ms":209,"total_runtime_ms":1182,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0165,"confidence_interval":0.9408,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.668,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1113,"p95_latency_ms":1360,"p99_latency_ms":1459,"time_to_first_token_ms":190,"total_runtime_ms":1236,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0249,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7971,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1145,"p95_latency_ms":1399,"p99_latency_ms":1501,"time_to_first_token_ms":212,"total_runtime_ms":1272,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2192,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1146,"p95_latency_ms":1401,"p99_latency_ms":1503,"time_to_first_token_ms":218,"total_runtime_ms":1274,"cold_start_penalty_ms":421,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0232,"confidence_interval":0.9468,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0206,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":756,"p95_latency_ms":924,"p99_latency_ms":991,"time_to_first_token_ms":178,"total_runtime_ms":840,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0194,"confidence_interval":0.9317,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7969,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":698,"p95_latency_ms":853,"p99_latency_ms":915,"time_to_first_token_ms":175,"total_runtime_ms":775,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.9305,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6486,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":784,"p95_latency_ms":958,"p99_latency_ms":1028,"time_to_first_token_ms":169,"total_runtime_ms":871,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0253,"confidence_interval":0.9391,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7824,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":781,"p95_latency_ms":955,"p99_latency_ms":1025,"time_to_first_token_ms":154,"total_runtime_ms":868,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0155,"confidence_interval":0.943,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3263,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":786,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":158,"total_runtime_ms":873,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.9436,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3533,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":689,"p95_latency_ms":842,"p99_latency_ms":903,"time_to_first_token_ms":163,"total_runtime_ms":765,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0169,"confidence_interval":0.9417,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.128,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":718,"p95_latency_ms":878,"p99_latency_ms":942,"time_to_first_token_ms":177,"total_runtime_ms":798,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0265,"confidence_interval":0.9333,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6196,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":687,"p95_latency_ms":840,"p99_latency_ms":901,"time_to_first_token_ms":176,"total_runtime_ms":763,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0256,"confidence_interval":0.9438,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6934,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":701,"p95_latency_ms":857,"p99_latency_ms":919,"time_to_first_token_ms":154,"total_runtime_ms":779,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0115,"confidence_interval":0.9308,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2754,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":783,"p95_latency_ms":957,"p99_latency_ms":1027,"time_to_first_token_ms":169,"total_runtime_ms":870,"cold_start_penalty_ms":0,"peak_ram_mb":1225,"avg_ram_mb":816,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0202,"confidence_interval":0.9477,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7434,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2221,"p95_latency_ms":2714,"p99_latency_ms":2911,"time_to_first_token_ms":340,"total_runtime_ms":2467,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.9401,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1299,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2137,"p95_latency_ms":2612,"p99_latency_ms":2802,"time_to_first_token_ms":337,"total_runtime_ms":2375,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0137,"confidence_interval":0.9394,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0338,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2230,"p95_latency_ms":2726,"p99_latency_ms":2924,"time_to_first_token_ms":349,"total_runtime_ms":2478,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0102,"confidence_interval":0.9423,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5271,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2194,"p95_latency_ms":2681,"p99_latency_ms":2876,"time_to_first_token_ms":336,"total_runtime_ms":2437,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.014,"confidence_interval":0.9475,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2481,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2144,"p95_latency_ms":2621,"p99_latency_ms":2812,"time_to_first_token_ms":325,"total_runtime_ms":2383,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0293,"confidence_interval":0.9384,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9431,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2136,"p95_latency_ms":2610,"p99_latency_ms":2800,"time_to_first_token_ms":352,"total_runtime_ms":2373,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0176,"confidence_interval":0.9422,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6605,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2207,"p95_latency_ms":2697,"p99_latency_ms":2893,"time_to_first_token_ms":335,"total_runtime_ms":2452,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.016,"confidence_interval":0.9364,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5031,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2159,"p95_latency_ms":2639,"p99_latency_ms":2831,"time_to_first_token_ms":345,"total_runtime_ms":2399,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0293,"confidence_interval":0.9378,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1686,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2227,"p95_latency_ms":2721,"p99_latency_ms":2919,"time_to_first_token_ms":347,"total_runtime_ms":2474,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0156,"confidence_interval":0.937,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.3757,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2127,"p95_latency_ms":2599,"p99_latency_ms":2788,"time_to_first_token_ms":340,"total_runtime_ms":2363,"cold_start_penalty_ms":842,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0252,"confidence_interval":0.9343,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.904,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1447,"p95_latency_ms":1769,"p99_latency_ms":1897,"time_to_first_token_ms":284,"total_runtime_ms":1608,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9315,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5973,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1370,"p95_latency_ms":1674,"p99_latency_ms":1796,"time_to_first_token_ms":290,"total_runtime_ms":1522,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0256,"confidence_interval":0.9431,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3021,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1455,"p95_latency_ms":1778,"p99_latency_ms":1908,"time_to_first_token_ms":302,"total_runtime_ms":1617,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0224,"confidence_interval":0.9421,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4491,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1445,"p95_latency_ms":1767,"p99_latency_ms":1895,"time_to_first_token_ms":294,"total_runtime_ms":1606,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.012,"confidence_interval":0.9373,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5503,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1435,"p95_latency_ms":1754,"p99_latency_ms":1882,"time_to_first_token_ms":278,"total_runtime_ms":1595,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0134,"confidence_interval":0.9443,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6212,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1445,"p95_latency_ms":1766,"p99_latency_ms":1894,"time_to_first_token_ms":281,"total_runtime_ms":1605,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.026,"confidence_interval":0.9464,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7683,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1399,"p95_latency_ms":1710,"p99_latency_ms":1834,"time_to_first_token_ms":294,"total_runtime_ms":1554,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0192,"confidence_interval":0.931,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.0277,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1366,"p95_latency_ms":1670,"p99_latency_ms":1792,"time_to_first_token_ms":309,"total_runtime_ms":1518,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0108,"confidence_interval":0.9303,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7101,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1427,"p95_latency_ms":1745,"p99_latency_ms":1872,"time_to_first_token_ms":292,"total_runtime_ms":1586,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0153,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9167,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1471,"p95_latency_ms":1798,"p99_latency_ms":1929,"time_to_first_token_ms":297,"total_runtime_ms":1634,"cold_start_penalty_ms":0,"peak_ram_mb":2449,"avg_ram_mb":1633,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0142,"confidence_interval":0.9474,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.3147,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3248,"p95_latency_ms":3970,"p99_latency_ms":4259,"time_to_first_token_ms":451,"total_runtime_ms":3609,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0237,"confidence_interval":0.9399,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2258,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3169,"p95_latency_ms":3873,"p99_latency_ms":4155,"time_to_first_token_ms":465,"total_runtime_ms":3521,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.01,"confidence_interval":0.9304,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.9526,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3209,"p95_latency_ms":3922,"p99_latency_ms":4207,"time_to_first_token_ms":460,"total_runtime_ms":3565,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0116,"confidence_interval":0.9323,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.849,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3157,"p95_latency_ms":3858,"p99_latency_ms":4139,"time_to_first_token_ms":477,"total_runtime_ms":3507,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0203,"confidence_interval":0.9326,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5033,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3253,"p95_latency_ms":3975,"p99_latency_ms":4264,"time_to_first_token_ms":476,"total_runtime_ms":3614,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0118,"confidence_interval":0.9429,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3763,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3250,"p95_latency_ms":3973,"p99_latency_ms":4261,"time_to_first_token_ms":446,"total_runtime_ms":3611,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0263,"confidence_interval":0.9313,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5693,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3149,"p95_latency_ms":3848,"p99_latency_ms":4128,"time_to_first_token_ms":450,"total_runtime_ms":3498,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0257,"confidence_interval":0.9496,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6019,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3202,"p95_latency_ms":3913,"p99_latency_ms":4198,"time_to_first_token_ms":476,"total_runtime_ms":3557,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0155,"confidence_interval":0.93,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4919,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3222,"p95_latency_ms":3938,"p99_latency_ms":4225,"time_to_first_token_ms":451,"total_runtime_ms":3580,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0215,"confidence_interval":0.9306,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1974,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":3167,"p95_latency_ms":3870,"p99_latency_ms":4152,"time_to_first_token_ms":451,"total_runtime_ms":3519,"cold_start_penalty_ms":1248,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0221,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4269,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2087,"p95_latency_ms":2551,"p99_latency_ms":2737,"time_to_first_token_ms":413,"total_runtime_ms":2319,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.9381,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.474,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2040,"p95_latency_ms":2493,"p99_latency_ms":2674,"time_to_first_token_ms":418,"total_runtime_ms":2266,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0119,"confidence_interval":0.9475,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1718,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2121,"p95_latency_ms":2593,"p99_latency_ms":2781,"time_to_first_token_ms":415,"total_runtime_ms":2357,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.02,"confidence_interval":0.9352,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6013,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2036,"p95_latency_ms":2488,"p99_latency_ms":2669,"time_to_first_token_ms":401,"total_runtime_ms":2262,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0238,"confidence_interval":0.9422,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1814,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2092,"p95_latency_ms":2557,"p99_latency_ms":2743,"time_to_first_token_ms":429,"total_runtime_ms":2325,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0222,"confidence_interval":0.937,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4969,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2096,"p95_latency_ms":2561,"p99_latency_ms":2748,"time_to_first_token_ms":420,"total_runtime_ms":2329,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0236,"confidence_interval":0.934,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1185,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2056,"p95_latency_ms":2513,"p99_latency_ms":2696,"time_to_first_token_ms":427,"total_runtime_ms":2284,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0195,"confidence_interval":0.938,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.6314,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2118,"p95_latency_ms":2589,"p99_latency_ms":2778,"time_to_first_token_ms":403,"total_runtime_ms":2354,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0277,"confidence_interval":0.9458,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3181,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2125,"p95_latency_ms":2597,"p99_latency_ms":2786,"time_to_first_token_ms":413,"total_runtime_ms":2361,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0249,"confidence_interval":0.9321,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.3875,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":2027,"p95_latency_ms":2477,"p99_latency_ms":2657,"time_to_first_token_ms":408,"total_runtime_ms":2252,"cold_start_penalty_ms":0,"peak_ram_mb":3629,"avg_ram_mb":2419,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0106,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4927,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":590,"p95_latency_ms":721,"p99_latency_ms":774,"time_to_first_token_ms":165,"total_runtime_ms":656,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0123,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8341,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":646,"p95_latency_ms":789,"p99_latency_ms":847,"time_to_first_token_ms":142,"total_runtime_ms":718,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7032,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":621,"p95_latency_ms":759,"p99_latency_ms":815,"time_to_first_token_ms":164,"total_runtime_ms":690,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0252,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4851,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":660,"p95_latency_ms":806,"p99_latency_ms":865,"time_to_first_token_ms":136,"total_runtime_ms":733,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0125,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3528,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":697,"p95_latency_ms":852,"p99_latency_ms":914,"time_to_first_token_ms":161,"total_runtime_ms":775,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0284,"confidence_interval":0.9438,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7274,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":615,"p95_latency_ms":751,"p99_latency_ms":806,"time_to_first_token_ms":157,"total_runtime_ms":683,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0204,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.197,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":630,"p95_latency_ms":770,"p99_latency_ms":826,"time_to_first_token_ms":137,"total_runtime_ms":700,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0234,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9126,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":611,"p95_latency_ms":747,"p99_latency_ms":802,"time_to_first_token_ms":139,"total_runtime_ms":679,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.016,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3148,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":653,"p95_latency_ms":798,"p99_latency_ms":856,"time_to_first_token_ms":149,"total_runtime_ms":725,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0272,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9513,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":595,"p95_latency_ms":727,"p99_latency_ms":780,"time_to_first_token_ms":154,"total_runtime_ms":661,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0181,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7005,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":461,"p95_latency_ms":564,"p99_latency_ms":605,"time_to_first_token_ms":113,"total_runtime_ms":512,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0274,"confidence_interval":0.9307,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4614,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":507,"time_to_first_token_ms":107,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0103,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9149,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":592,"time_to_first_token_ms":104,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0285,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6724,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":460,"p95_latency_ms":563,"p99_latency_ms":604,"time_to_first_token_ms":105,"total_runtime_ms":511,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0146,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3513,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":537,"time_to_first_token_ms":94,"total_runtime_ms":455,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0268,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0105,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":480,"p95_latency_ms":586,"p99_latency_ms":629,"time_to_first_token_ms":114,"total_runtime_ms":533,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0166,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0236,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":420,"p95_latency_ms":514,"p99_latency_ms":551,"time_to_first_token_ms":116,"total_runtime_ms":467,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0147,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7792,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":474,"p95_latency_ms":579,"p99_latency_ms":621,"time_to_first_token_ms":94,"total_runtime_ms":526,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.015,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1886,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":469,"p95_latency_ms":573,"p99_latency_ms":615,"time_to_first_token_ms":103,"total_runtime_ms":521,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0228,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8438,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":111,"total_runtime_ms":436,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0191,"confidence_interval":0.9381,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.661,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":889,"p95_latency_ms":1086,"p99_latency_ms":1165,"time_to_first_token_ms":174,"total_runtime_ms":987,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0114,"confidence_interval":0.9424,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0127,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":878,"p95_latency_ms":1073,"p99_latency_ms":1151,"time_to_first_token_ms":181,"total_runtime_ms":976,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9386,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5972,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":806,"p95_latency_ms":986,"p99_latency_ms":1057,"time_to_first_token_ms":176,"total_runtime_ms":896,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0279,"confidence_interval":0.9332,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3006,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":845,"p95_latency_ms":1033,"p99_latency_ms":1108,"time_to_first_token_ms":192,"total_runtime_ms":939,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.9395,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2515,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":840,"p95_latency_ms":1027,"p99_latency_ms":1101,"time_to_first_token_ms":186,"total_runtime_ms":933,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0131,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2397,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":862,"p95_latency_ms":1053,"p99_latency_ms":1130,"time_to_first_token_ms":174,"total_runtime_ms":958,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0242,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7681,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":839,"p95_latency_ms":1025,"p99_latency_ms":1099,"time_to_first_token_ms":165,"total_runtime_ms":932,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0113,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8528,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":833,"p95_latency_ms":1018,"p99_latency_ms":1092,"time_to_first_token_ms":168,"total_runtime_ms":925,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0103,"confidence_interval":0.9326,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6477,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":833,"p95_latency_ms":1018,"p99_latency_ms":1092,"time_to_first_token_ms":159,"total_runtime_ms":926,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0265,"confidence_interval":0.9324,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3189,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":865,"p95_latency_ms":1057,"p99_latency_ms":1134,"time_to_first_token_ms":176,"total_runtime_ms":961,"cold_start_penalty_ms":316,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":62,"gpu_utilization_percent":49,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0109,"confidence_interval":0.9458,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4562,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":756,"time_to_first_token_ms":115,"total_runtime_ms":641,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0191,"confidence_interval":0.9362,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1669,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":592,"p95_latency_ms":724,"p99_latency_ms":776,"time_to_first_token_ms":141,"total_runtime_ms":658,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0168,"confidence_interval":0.9389,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8609,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":581,"p95_latency_ms":710,"p99_latency_ms":762,"time_to_first_token_ms":140,"total_runtime_ms":646,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0141,"confidence_interval":0.9477,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5717,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":613,"p95_latency_ms":749,"p99_latency_ms":803,"time_to_first_token_ms":119,"total_runtime_ms":681,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0253,"confidence_interval":0.9468,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1148,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":544,"p95_latency_ms":665,"p99_latency_ms":713,"time_to_first_token_ms":144,"total_runtime_ms":604,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0286,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2159,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":556,"p95_latency_ms":680,"p99_latency_ms":729,"time_to_first_token_ms":147,"total_runtime_ms":618,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0228,"confidence_interval":0.9345,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3413,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":583,"p95_latency_ms":713,"p99_latency_ms":764,"time_to_first_token_ms":123,"total_runtime_ms":648,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0179,"confidence_interval":0.9464,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3912,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":138,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0188,"confidence_interval":0.9499,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.802,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":569,"p95_latency_ms":696,"p99_latency_ms":746,"time_to_first_token_ms":132,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0294,"confidence_interval":0.949,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5045,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":773,"time_to_first_token_ms":131,"total_runtime_ms":655,"cold_start_penalty_ms":0,"peak_ram_mb":919,"avg_ram_mb":612,"cpu_utilization_percent":52,"gpu_utilization_percent":49,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0277,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2939,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1645,"p95_latency_ms":2011,"p99_latency_ms":2157,"time_to_first_token_ms":283,"total_runtime_ms":1828,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0257,"confidence_interval":0.9395,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4792,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1635,"p95_latency_ms":1999,"p99_latency_ms":2144,"time_to_first_token_ms":273,"total_runtime_ms":1817,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0198,"confidence_interval":0.941,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6253,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1642,"p95_latency_ms":2006,"p99_latency_ms":2152,"time_to_first_token_ms":265,"total_runtime_ms":1824,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0292,"confidence_interval":0.9348,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0514,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1684,"p95_latency_ms":2059,"p99_latency_ms":2209,"time_to_first_token_ms":288,"total_runtime_ms":1872,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0145,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.3205,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1638,"p95_latency_ms":2002,"p99_latency_ms":2148,"time_to_first_token_ms":274,"total_runtime_ms":1820,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0178,"confidence_interval":0.9307,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0195,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1605,"p95_latency_ms":1961,"p99_latency_ms":2104,"time_to_first_token_ms":277,"total_runtime_ms":1783,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0254,"confidence_interval":0.9477,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5066,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1686,"p95_latency_ms":2060,"p99_latency_ms":2210,"time_to_first_token_ms":269,"total_runtime_ms":1873,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0137,"confidence_interval":0.9377,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4926,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1689,"p95_latency_ms":2064,"p99_latency_ms":2214,"time_to_first_token_ms":273,"total_runtime_ms":1876,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0102,"confidence_interval":0.947,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.3027,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1660,"p95_latency_ms":2029,"p99_latency_ms":2177,"time_to_first_token_ms":286,"total_runtime_ms":1845,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.01,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0129,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1662,"p95_latency_ms":2032,"p99_latency_ms":2179,"time_to_first_token_ms":268,"total_runtime_ms":1847,"cold_start_penalty_ms":632,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":62,"gpu_utilization_percent":97,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0243,"confidence_interval":0.9397,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7783,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1130,"p95_latency_ms":1381,"p99_latency_ms":1481,"time_to_first_token_ms":234,"total_runtime_ms":1255,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0224,"confidence_interval":0.9424,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5073,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1117,"p95_latency_ms":1366,"p99_latency_ms":1465,"time_to_first_token_ms":228,"total_runtime_ms":1241,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.938,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.513,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1045,"p95_latency_ms":1277,"p99_latency_ms":1370,"time_to_first_token_ms":210,"total_runtime_ms":1161,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0232,"confidence_interval":0.9369,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.4024,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1048,"p95_latency_ms":1281,"p99_latency_ms":1374,"time_to_first_token_ms":239,"total_runtime_ms":1164,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0115,"confidence_interval":0.9337,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.628,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1071,"p95_latency_ms":1310,"p99_latency_ms":1405,"time_to_first_token_ms":244,"total_runtime_ms":1190,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0207,"confidence_interval":0.9372,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1959,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1025,"p95_latency_ms":1252,"p99_latency_ms":1343,"time_to_first_token_ms":234,"total_runtime_ms":1139,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.017,"confidence_interval":0.9471,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0552,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1110,"p95_latency_ms":1357,"p99_latency_ms":1456,"time_to_first_token_ms":215,"total_runtime_ms":1234,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0203,"confidence_interval":0.9311,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2467,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1104,"p95_latency_ms":1349,"p99_latency_ms":1447,"time_to_first_token_ms":210,"total_runtime_ms":1227,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0173,"confidence_interval":0.9356,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2455,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1045,"p95_latency_ms":1277,"p99_latency_ms":1370,"time_to_first_token_ms":212,"total_runtime_ms":1161,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0172,"confidence_interval":0.9488,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.504,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1103,"p95_latency_ms":1348,"p99_latency_ms":1446,"time_to_first_token_ms":242,"total_runtime_ms":1225,"cold_start_penalty_ms":0,"peak_ram_mb":1837,"avg_ram_mb":1225,"cpu_utilization_percent":52,"gpu_utilization_percent":97,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.013,"confidence_interval":0.9486,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8633,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2368,"p95_latency_ms":2895,"p99_latency_ms":3105,"time_to_first_token_ms":362,"total_runtime_ms":2631,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0208,"confidence_interval":0.9392,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1776,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2404,"p95_latency_ms":2938,"p99_latency_ms":3151,"time_to_first_token_ms":348,"total_runtime_ms":2671,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0203,"confidence_interval":0.9323,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.027,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2408,"p95_latency_ms":2944,"p99_latency_ms":3158,"time_to_first_token_ms":381,"total_runtime_ms":2676,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0132,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3295,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2398,"p95_latency_ms":2931,"p99_latency_ms":3144,"time_to_first_token_ms":371,"total_runtime_ms":2664,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0235,"confidence_interval":0.9431,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0723,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2461,"p95_latency_ms":3008,"p99_latency_ms":3226,"time_to_first_token_ms":348,"total_runtime_ms":2734,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0142,"confidence_interval":0.9406,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1996,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2464,"p95_latency_ms":3011,"p99_latency_ms":3230,"time_to_first_token_ms":370,"total_runtime_ms":2737,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0259,"confidence_interval":0.9418,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9027,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2444,"p95_latency_ms":2987,"p99_latency_ms":3204,"time_to_first_token_ms":365,"total_runtime_ms":2716,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0126,"confidence_interval":0.9388,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9931,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2365,"p95_latency_ms":2890,"p99_latency_ms":3101,"time_to_first_token_ms":352,"total_runtime_ms":2628,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0215,"confidence_interval":0.9363,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4357,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2405,"p95_latency_ms":2940,"p99_latency_ms":3153,"time_to_first_token_ms":355,"total_runtime_ms":2672,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0104,"confidence_interval":0.9433,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3494,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2447,"p95_latency_ms":2991,"p99_latency_ms":3208,"time_to_first_token_ms":350,"total_runtime_ms":2719,"cold_start_penalty_ms":936,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":62,"gpu_utilization_percent":100,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0213,"confidence_interval":0.9396,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4249,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1576,"p95_latency_ms":1926,"p99_latency_ms":2067,"time_to_first_token_ms":307,"total_runtime_ms":1751,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0204,"confidence_interval":0.9364,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2592,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1569,"p95_latency_ms":1918,"p99_latency_ms":2057,"time_to_first_token_ms":326,"total_runtime_ms":1743,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0152,"confidence_interval":0.9425,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.798,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1624,"p95_latency_ms":1984,"p99_latency_ms":2129,"time_to_first_token_ms":325,"total_runtime_ms":1804,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0159,"confidence_interval":0.9428,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5328,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1523,"p95_latency_ms":1862,"p99_latency_ms":1997,"time_to_first_token_ms":322,"total_runtime_ms":1692,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0208,"confidence_interval":0.9473,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5318,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1539,"p95_latency_ms":1881,"p99_latency_ms":2018,"time_to_first_token_ms":305,"total_runtime_ms":1710,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0184,"confidence_interval":0.9314,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9858,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1535,"p95_latency_ms":1876,"p99_latency_ms":2013,"time_to_first_token_ms":330,"total_runtime_ms":1706,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0209,"confidence_interval":0.9466,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8031,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1602,"p95_latency_ms":1958,"p99_latency_ms":2100,"time_to_first_token_ms":313,"total_runtime_ms":1780,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0145,"confidence_interval":0.9407,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1629,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1593,"p95_latency_ms":1946,"p99_latency_ms":2088,"time_to_first_token_ms":327,"total_runtime_ms":1770,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0187,"confidence_interval":0.9317,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2052,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1574,"p95_latency_ms":1924,"p99_latency_ms":2064,"time_to_first_token_ms":337,"total_runtime_ms":1749,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0202,"confidence_interval":0.9352,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3144,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1590,"p95_latency_ms":1943,"p99_latency_ms":2085,"time_to_first_token_ms":338,"total_runtime_ms":1767,"cold_start_penalty_ms":0,"peak_ram_mb":2722,"avg_ram_mb":1814,"cpu_utilization_percent":52,"gpu_utilization_percent":100,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0233,"confidence_interval":0.9327,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.6561,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":671,"time_to_first_token_ms":139,"total_runtime_ms":569,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0102,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4462,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":500,"p95_latency_ms":611,"p99_latency_ms":655,"time_to_first_token_ms":142,"total_runtime_ms":555,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0154,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.998,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":574,"p95_latency_ms":702,"p99_latency_ms":753,"time_to_first_token_ms":147,"total_runtime_ms":638,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0104,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1635,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":585,"p95_latency_ms":716,"p99_latency_ms":768,"time_to_first_token_ms":141,"total_runtime_ms":650,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0188,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3143,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":525,"p95_latency_ms":642,"p99_latency_ms":688,"time_to_first_token_ms":148,"total_runtime_ms":583,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0138,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3399,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":668,"time_to_first_token_ms":129,"total_runtime_ms":566,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0278,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1885,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":508,"p95_latency_ms":621,"p99_latency_ms":666,"time_to_first_token_ms":152,"total_runtime_ms":565,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0183,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9593,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":543,"p95_latency_ms":664,"p99_latency_ms":713,"time_to_first_token_ms":136,"total_runtime_ms":604,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0156,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3728,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":590,"p95_latency_ms":721,"p99_latency_ms":773,"time_to_first_token_ms":128,"total_runtime_ms":655,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.014,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3459,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":504,"p95_latency_ms":616,"p99_latency_ms":661,"time_to_first_token_ms":128,"total_runtime_ms":560,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0273,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4164,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":366,"p95_latency_ms":447,"p99_latency_ms":480,"time_to_first_token_ms":79,"total_runtime_ms":407,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9003,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":405,"p95_latency_ms":495,"p99_latency_ms":531,"time_to_first_token_ms":85,"total_runtime_ms":450,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0273,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7496,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":362,"p95_latency_ms":442,"p99_latency_ms":474,"time_to_first_token_ms":94,"total_runtime_ms":402,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.014,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.091,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":367,"p95_latency_ms":449,"p99_latency_ms":482,"time_to_first_token_ms":102,"total_runtime_ms":408,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0276,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.136,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":477,"time_to_first_token_ms":86,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0292,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6394,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":100,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0134,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2066,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":331,"p95_latency_ms":404,"p99_latency_ms":434,"time_to_first_token_ms":95,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.024,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9559,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":394,"p95_latency_ms":482,"p99_latency_ms":517,"time_to_first_token_ms":90,"total_runtime_ms":438,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0267,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1155,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":333,"p95_latency_ms":407,"p99_latency_ms":437,"time_to_first_token_ms":91,"total_runtime_ms":370,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0228,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7336,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":506,"time_to_first_token_ms":108,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.012,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4581,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":762,"p95_latency_ms":931,"p99_latency_ms":998,"time_to_first_token_ms":144,"total_runtime_ms":846,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0298,"confidence_interval":0.9301,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0779,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1009,"time_to_first_token_ms":142,"total_runtime_ms":855,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0266,"confidence_interval":0.9461,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8848,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":768,"p95_latency_ms":939,"p99_latency_ms":1007,"time_to_first_token_ms":175,"total_runtime_ms":854,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0262,"confidence_interval":0.946,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.061,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":160,"total_runtime_ms":764,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0247,"confidence_interval":0.9318,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4898,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":666,"p95_latency_ms":814,"p99_latency_ms":873,"time_to_first_token_ms":141,"total_runtime_ms":740,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0158,"confidence_interval":0.9432,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7755,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":688,"p95_latency_ms":840,"p99_latency_ms":902,"time_to_first_token_ms":170,"total_runtime_ms":764,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0228,"confidence_interval":0.938,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9063,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":708,"p95_latency_ms":866,"p99_latency_ms":929,"time_to_first_token_ms":166,"total_runtime_ms":787,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0263,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":758,"p95_latency_ms":927,"p99_latency_ms":994,"time_to_first_token_ms":142,"total_runtime_ms":843,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0191,"confidence_interval":0.9374,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4395,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":712,"p95_latency_ms":870,"p99_latency_ms":933,"time_to_first_token_ms":166,"total_runtime_ms":791,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.469,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":739,"p95_latency_ms":903,"p99_latency_ms":969,"time_to_first_token_ms":157,"total_runtime_ms":821,"cold_start_penalty_ms":263,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0287,"confidence_interval":0.9405,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8434,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":466,"p95_latency_ms":570,"p99_latency_ms":611,"time_to_first_token_ms":103,"total_runtime_ms":518,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0294,"confidence_interval":0.9453,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0831,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":112,"total_runtime_ms":504,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.019,"confidence_interval":0.9301,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9704,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":467,"p95_latency_ms":571,"p99_latency_ms":613,"time_to_first_token_ms":115,"total_runtime_ms":519,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0185,"confidence_interval":0.9428,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1069,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":468,"p95_latency_ms":572,"p99_latency_ms":614,"time_to_first_token_ms":113,"total_runtime_ms":520,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0178,"confidence_interval":0.9344,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1297,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":534,"p95_latency_ms":653,"p99_latency_ms":700,"time_to_first_token_ms":99,"total_runtime_ms":593,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0233,"confidence_interval":0.9499,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7993,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":434,"p95_latency_ms":531,"p99_latency_ms":569,"time_to_first_token_ms":110,"total_runtime_ms":482,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.027,"confidence_interval":0.9432,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3242,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":441,"p95_latency_ms":539,"p99_latency_ms":578,"time_to_first_token_ms":110,"total_runtime_ms":490,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9488,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3439,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":456,"p95_latency_ms":557,"p99_latency_ms":598,"time_to_first_token_ms":127,"total_runtime_ms":507,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0283,"confidence_interval":0.9343,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8798,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":478,"p95_latency_ms":584,"p99_latency_ms":626,"time_to_first_token_ms":123,"total_runtime_ms":531,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0121,"confidence_interval":0.9346,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9621,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":498,"p95_latency_ms":608,"p99_latency_ms":653,"time_to_first_token_ms":119,"total_runtime_ms":553,"cold_start_penalty_ms":0,"peak_ram_mb":765,"avg_ram_mb":510,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0151,"confidence_interval":0.9425,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2465,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1378,"p95_latency_ms":1684,"p99_latency_ms":1807,"time_to_first_token_ms":228,"total_runtime_ms":1531,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0132,"confidence_interval":0.9391,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4633,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1369,"p95_latency_ms":1673,"p99_latency_ms":1794,"time_to_first_token_ms":248,"total_runtime_ms":1521,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0168,"confidence_interval":0.9481,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8523,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1343,"p95_latency_ms":1642,"p99_latency_ms":1761,"time_to_first_token_ms":233,"total_runtime_ms":1492,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0268,"confidence_interval":0.9438,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2855,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1406,"p95_latency_ms":1718,"p99_latency_ms":1843,"time_to_first_token_ms":247,"total_runtime_ms":1562,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0267,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8507,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1359,"p95_latency_ms":1661,"p99_latency_ms":1782,"time_to_first_token_ms":237,"total_runtime_ms":1510,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0168,"confidence_interval":0.9382,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3922,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1390,"p95_latency_ms":1698,"p99_latency_ms":1822,"time_to_first_token_ms":238,"total_runtime_ms":1544,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0236,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8676,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1346,"p95_latency_ms":1645,"p99_latency_ms":1764,"time_to_first_token_ms":225,"total_runtime_ms":1495,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0203,"confidence_interval":0.9306,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8274,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1434,"p95_latency_ms":1753,"p99_latency_ms":1881,"time_to_first_token_ms":226,"total_runtime_ms":1594,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0187,"confidence_interval":0.9442,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0034,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1354,"p95_latency_ms":1655,"p99_latency_ms":1775,"time_to_first_token_ms":223,"total_runtime_ms":1504,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0283,"confidence_interval":0.9495,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.5789,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1339,"p95_latency_ms":1637,"p99_latency_ms":1756,"time_to_first_token_ms":250,"total_runtime_ms":1488,"cold_start_penalty_ms":527,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0172,"confidence_interval":0.9311,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.3084,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":941,"p95_latency_ms":1150,"p99_latency_ms":1234,"time_to_first_token_ms":179,"total_runtime_ms":1046,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0297,"confidence_interval":0.9387,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6153,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":934,"p95_latency_ms":1142,"p99_latency_ms":1225,"time_to_first_token_ms":187,"total_runtime_ms":1038,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0164,"confidence_interval":0.9474,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2642,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":861,"p95_latency_ms":1053,"p99_latency_ms":1129,"time_to_first_token_ms":198,"total_runtime_ms":957,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0259,"confidence_interval":0.9478,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0861,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":960,"p95_latency_ms":1174,"p99_latency_ms":1259,"time_to_first_token_ms":208,"total_runtime_ms":1067,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0111,"confidence_interval":0.9446,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9474,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":947,"p95_latency_ms":1157,"p99_latency_ms":1241,"time_to_first_token_ms":192,"total_runtime_ms":1052,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0178,"confidence_interval":0.9416,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9572,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":853,"p95_latency_ms":1043,"p99_latency_ms":1118,"time_to_first_token_ms":179,"total_runtime_ms":948,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0277,"confidence_interval":0.9432,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4265,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":900,"p95_latency_ms":1100,"p99_latency_ms":1180,"time_to_first_token_ms":196,"total_runtime_ms":1000,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.011,"confidence_interval":0.9371,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1739,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":952,"p95_latency_ms":1163,"p99_latency_ms":1248,"time_to_first_token_ms":186,"total_runtime_ms":1057,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0154,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7792,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":946,"p95_latency_ms":1157,"p99_latency_ms":1241,"time_to_first_token_ms":184,"total_runtime_ms":1051,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0102,"confidence_interval":0.9396,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1018,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:7b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":894,"p95_latency_ms":1093,"p99_latency_ms":1173,"time_to_first_token_ms":210,"total_runtime_ms":994,"cold_start_penalty_ms":0,"peak_ram_mb":1531,"avg_ram_mb":1021,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0205,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2279,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2046,"p95_latency_ms":2501,"p99_latency_ms":2683,"time_to_first_token_ms":329,"total_runtime_ms":2274,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.9499,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2986,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1994,"p95_latency_ms":2438,"p99_latency_ms":2615,"time_to_first_token_ms":304,"total_runtime_ms":2216,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0297,"confidence_interval":0.9472,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1745,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2018,"p95_latency_ms":2466,"p99_latency_ms":2646,"time_to_first_token_ms":301,"total_runtime_ms":2242,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0171,"confidence_interval":0.9366,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9425,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2073,"p95_latency_ms":2533,"p99_latency_ms":2718,"time_to_first_token_ms":318,"total_runtime_ms":2303,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0253,"confidence_interval":0.9499,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4671,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2002,"p95_latency_ms":2447,"p99_latency_ms":2625,"time_to_first_token_ms":319,"total_runtime_ms":2225,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0221,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5328,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2045,"p95_latency_ms":2499,"p99_latency_ms":2681,"time_to_first_token_ms":330,"total_runtime_ms":2272,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0257,"confidence_interval":0.9329,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9776,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2010,"p95_latency_ms":2457,"p99_latency_ms":2636,"time_to_first_token_ms":317,"total_runtime_ms":2234,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0178,"confidence_interval":0.9318,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4147,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2011,"p95_latency_ms":2458,"p99_latency_ms":2637,"time_to_first_token_ms":302,"total_runtime_ms":2234,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.948,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.127,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":2040,"p95_latency_ms":2493,"p99_latency_ms":2674,"time_to_first_token_ms":307,"total_runtime_ms":2266,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0288,"confidence_interval":0.9305,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0428,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.9,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1993,"p95_latency_ms":2436,"p99_latency_ms":2613,"time_to_first_token_ms":319,"total_runtime_ms":2214,"cold_start_penalty_ms":780,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0249,"confidence_interval":0.9481,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7645,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1276,"p95_latency_ms":1560,"p99_latency_ms":1674,"time_to_first_token_ms":271,"total_runtime_ms":1418,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0272,"confidence_interval":0.9357,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0089,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1359,"p95_latency_ms":1661,"p99_latency_ms":1781,"time_to_first_token_ms":273,"total_runtime_ms":1510,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0171,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8267,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1290,"p95_latency_ms":1577,"p99_latency_ms":1691,"time_to_first_token_ms":257,"total_runtime_ms":1433,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0144,"confidence_interval":0.9357,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0558,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1360,"p95_latency_ms":1663,"p99_latency_ms":1784,"time_to_first_token_ms":259,"total_runtime_ms":1512,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.017,"confidence_interval":0.9439,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6554,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1297,"p95_latency_ms":1586,"p99_latency_ms":1701,"time_to_first_token_ms":282,"total_runtime_ms":1442,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0149,"confidence_interval":0.9422,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5157,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1343,"p95_latency_ms":1642,"p99_latency_ms":1761,"time_to_first_token_ms":257,"total_runtime_ms":1493,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0165,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.589,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1290,"p95_latency_ms":1577,"p99_latency_ms":1691,"time_to_first_token_ms":279,"total_runtime_ms":1433,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.027,"confidence_interval":0.939,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.317,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1371,"p95_latency_ms":1676,"p99_latency_ms":1798,"time_to_first_token_ms":281,"total_runtime_ms":1524,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0103,"confidence_interval":0.9408,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7184,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1364,"p95_latency_ms":1667,"p99_latency_ms":1789,"time_to_first_token_ms":260,"total_runtime_ms":1516,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.01,"confidence_interval":0.9487,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.359,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"long_context|high|qwen2:14b|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1284,"p95_latency_ms":1569,"p99_latency_ms":1683,"time_to_first_token_ms":285,"total_runtime_ms":1427,"cold_start_penalty_ms":0,"peak_ram_mb":2268,"avg_ram_mb":1512,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.018,"confidence_interval":0.9492,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.025,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":755,"p95_latency_ms":923,"p99_latency_ms":990,"time_to_first_token_ms":189,"total_runtime_ms":839,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0191,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3305,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":778,"p95_latency_ms":951,"p99_latency_ms":1020,"time_to_first_token_ms":192,"total_runtime_ms":865,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0283,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9212,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":848,"p95_latency_ms":1036,"p99_latency_ms":1112,"time_to_first_token_ms":222,"total_runtime_ms":942,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0213,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2787,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":749,"p95_latency_ms":916,"p99_latency_ms":982,"time_to_first_token_ms":213,"total_runtime_ms":833,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0277,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2461,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":796,"p95_latency_ms":973,"p99_latency_ms":1043,"time_to_first_token_ms":203,"total_runtime_ms":884,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0263,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.312,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":755,"p95_latency_ms":923,"p99_latency_ms":990,"time_to_first_token_ms":215,"total_runtime_ms":839,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.018,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2337,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":799,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":211,"total_runtime_ms":887,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0166,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2803,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":767,"p95_latency_ms":938,"p99_latency_ms":1006,"time_to_first_token_ms":212,"total_runtime_ms":852,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0221,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3935,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":786,"p95_latency_ms":960,"p99_latency_ms":1030,"time_to_first_token_ms":198,"total_runtime_ms":873,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0287,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9648,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":825,"p95_latency_ms":1008,"p99_latency_ms":1081,"time_to_first_token_ms":194,"total_runtime_ms":916,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0253,"confidence_interval":0.9347,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.397,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":470,"p95_latency_ms":575,"p99_latency_ms":617,"time_to_first_token_ms":164,"total_runtime_ms":523,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0276,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3893,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":444,"p95_latency_ms":542,"p99_latency_ms":582,"time_to_first_token_ms":176,"total_runtime_ms":493,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0248,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7624,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":388,"p95_latency_ms":475,"p99_latency_ms":509,"time_to_first_token_ms":168,"total_runtime_ms":431,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0179,"confidence_interval":0.9346,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5847,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":413,"p95_latency_ms":504,"p99_latency_ms":541,"time_to_first_token_ms":151,"total_runtime_ms":459,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0293,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4364,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":416,"p95_latency_ms":509,"p99_latency_ms":546,"time_to_first_token_ms":173,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0244,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9999,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":419,"p95_latency_ms":512,"p99_latency_ms":550,"time_to_first_token_ms":154,"total_runtime_ms":466,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0116,"confidence_interval":0.9395,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6177,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":477,"p95_latency_ms":583,"p99_latency_ms":625,"time_to_first_token_ms":165,"total_runtime_ms":530,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0225,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4209,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":447,"p95_latency_ms":546,"p99_latency_ms":586,"time_to_first_token_ms":144,"total_runtime_ms":497,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0147,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8948,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":398,"p95_latency_ms":487,"p99_latency_ms":522,"time_to_first_token_ms":164,"total_runtime_ms":442,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0116,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9795,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:0.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":429,"p95_latency_ms":524,"p99_latency_ms":562,"time_to_first_token_ms":158,"total_runtime_ms":476,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0214,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.767,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1065,"p95_latency_ms":1302,"p99_latency_ms":1397,"time_to_first_token_ms":247,"total_runtime_ms":1184,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0134,"confidence_interval":0.9301,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2308,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1106,"p95_latency_ms":1352,"p99_latency_ms":1450,"time_to_first_token_ms":251,"total_runtime_ms":1229,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0151,"confidence_interval":0.9323,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8696,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1105,"p95_latency_ms":1350,"p99_latency_ms":1448,"time_to_first_token_ms":235,"total_runtime_ms":1227,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0298,"confidence_interval":0.9448,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4361,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1044,"p95_latency_ms":1276,"p99_latency_ms":1369,"time_to_first_token_ms":265,"total_runtime_ms":1160,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3645,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1023,"p95_latency_ms":1250,"p99_latency_ms":1341,"time_to_first_token_ms":253,"total_runtime_ms":1137,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1198,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1030,"p95_latency_ms":1259,"p99_latency_ms":1350,"time_to_first_token_ms":254,"total_runtime_ms":1144,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0269,"confidence_interval":0.9383,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7311,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1098,"p95_latency_ms":1342,"p99_latency_ms":1439,"time_to_first_token_ms":262,"total_runtime_ms":1220,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0109,"confidence_interval":0.9395,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1323,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1025,"p95_latency_ms":1252,"p99_latency_ms":1344,"time_to_first_token_ms":246,"total_runtime_ms":1139,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0451,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1056,"p95_latency_ms":1290,"p99_latency_ms":1384,"time_to_first_token_ms":235,"total_runtime_ms":1173,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0294,"confidence_interval":0.937,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3947,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1032,"p95_latency_ms":1262,"p99_latency_ms":1354,"time_to_first_token_ms":257,"total_runtime_ms":1147,"cold_start_penalty_ms":562,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0202,"confidence_interval":0.9384,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2363,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":612,"p95_latency_ms":747,"p99_latency_ms":802,"time_to_first_token_ms":211,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0101,"confidence_interval":0.9416,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9404,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":208,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0122,"confidence_interval":0.9422,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3469,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":592,"p95_latency_ms":724,"p99_latency_ms":776,"time_to_first_token_ms":216,"total_runtime_ms":658,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0281,"confidence_interval":0.942,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5045,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":195,"total_runtime_ms":650,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0114,"confidence_interval":0.9403,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7028,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":532,"p95_latency_ms":651,"p99_latency_ms":698,"time_to_first_token_ms":212,"total_runtime_ms":591,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0184,"confidence_interval":0.9496,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9938,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":561,"p95_latency_ms":686,"p99_latency_ms":736,"time_to_first_token_ms":210,"total_runtime_ms":624,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0197,"confidence_interval":0.9426,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.417,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":608,"p95_latency_ms":743,"p99_latency_ms":797,"time_to_first_token_ms":210,"total_runtime_ms":675,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0234,"confidence_interval":0.9321,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.458,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":581,"p95_latency_ms":711,"p99_latency_ms":762,"time_to_first_token_ms":201,"total_runtime_ms":646,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0164,"confidence_interval":0.9479,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1652,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":507,"p95_latency_ms":619,"p99_latency_ms":664,"time_to_first_token_ms":190,"total_runtime_ms":563,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9368,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1751,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:1.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":547,"p95_latency_ms":669,"p99_latency_ms":717,"time_to_first_token_ms":214,"total_runtime_ms":608,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":907,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0259,"confidence_interval":0.9441,"validation_overhead_ms":35,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8629,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2061,"p95_latency_ms":2519,"p99_latency_ms":2702,"time_to_first_token_ms":406,"total_runtime_ms":2290,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0231,"confidence_interval":0.9463,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3697,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2070,"p95_latency_ms":2530,"p99_latency_ms":2714,"time_to_first_token_ms":412,"total_runtime_ms":2300,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0154,"confidence_interval":0.936,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.3148,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2030,"p95_latency_ms":2481,"p99_latency_ms":2662,"time_to_first_token_ms":410,"total_runtime_ms":2256,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0159,"confidence_interval":0.9439,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7556,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2042,"p95_latency_ms":2496,"p99_latency_ms":2678,"time_to_first_token_ms":421,"total_runtime_ms":2269,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0131,"confidence_interval":0.9351,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.1157,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2098,"p95_latency_ms":2564,"p99_latency_ms":2750,"time_to_first_token_ms":414,"total_runtime_ms":2331,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0258,"confidence_interval":0.9337,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.7129,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2094,"p95_latency_ms":2560,"p99_latency_ms":2746,"time_to_first_token_ms":418,"total_runtime_ms":2327,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0226,"confidence_interval":0.9398,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.0459,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2072,"p95_latency_ms":2532,"p99_latency_ms":2716,"time_to_first_token_ms":412,"total_runtime_ms":2302,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0164,"confidence_interval":0.9407,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5885,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2032,"p95_latency_ms":2484,"p99_latency_ms":2664,"time_to_first_token_ms":435,"total_runtime_ms":2258,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0251,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.498,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2097,"p95_latency_ms":2563,"p99_latency_ms":2749,"time_to_first_token_ms":432,"total_runtime_ms":2330,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.02,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4866,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":2091,"p95_latency_ms":2555,"p99_latency_ms":2741,"time_to_first_token_ms":440,"total_runtime_ms":2323,"cold_start_penalty_ms":1123,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0158,"confidence_interval":0.9458,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.475,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1013,"p95_latency_ms":1238,"p99_latency_ms":1329,"time_to_first_token_ms":384,"total_runtime_ms":1126,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9454,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3871,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1061,"p95_latency_ms":1296,"p99_latency_ms":1391,"time_to_first_token_ms":368,"total_runtime_ms":1179,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0117,"confidence_interval":0.942,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9571,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1034,"p95_latency_ms":1263,"p99_latency_ms":1355,"time_to_first_token_ms":384,"total_runtime_ms":1148,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0178,"confidence_interval":0.9441,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3881,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1112,"p95_latency_ms":1360,"p99_latency_ms":1458,"time_to_first_token_ms":389,"total_runtime_ms":1236,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0187,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6982,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1044,"p95_latency_ms":1276,"p99_latency_ms":1369,"time_to_first_token_ms":383,"total_runtime_ms":1160,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0285,"confidence_interval":0.9451,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.5067,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1048,"p95_latency_ms":1280,"p99_latency_ms":1374,"time_to_first_token_ms":362,"total_runtime_ms":1164,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.015,"confidence_interval":0.9302,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":9.3169,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1060,"p95_latency_ms":1295,"p99_latency_ms":1390,"time_to_first_token_ms":377,"total_runtime_ms":1178,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0177,"confidence_interval":0.947,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.3712,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1045,"p95_latency_ms":1277,"p99_latency_ms":1370,"time_to_first_token_ms":382,"total_runtime_ms":1161,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0123,"confidence_interval":0.9311,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.4101,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1029,"p95_latency_ms":1258,"p99_latency_ms":1350,"time_to_first_token_ms":364,"total_runtime_ms":1144,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0171,"confidence_interval":0.9418,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.9013,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:7b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":1051,"p95_latency_ms":1284,"p99_latency_ms":1378,"time_to_first_token_ms":367,"total_runtime_ms":1167,"cold_start_penalty_ms":0,"peak_ram_mb":2419,"avg_ram_mb":1814,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0187,"confidence_interval":0.9363,"validation_overhead_ms":46,"orchestration_overhead_ms":72,"event_dispatch_cost":8.6008,"memory_management_cost":9.84,"memory_pressure_events":0} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3036,"p95_latency_ms":3711,"p99_latency_ms":3981,"time_to_first_token_ms":605,"total_runtime_ms":3374,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0292,"confidence_interval":0.9334,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7848,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3020,"p95_latency_ms":3691,"p99_latency_ms":3960,"time_to_first_token_ms":599,"total_runtime_ms":3356,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0175,"confidence_interval":0.9459,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4157,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3062,"p95_latency_ms":3742,"p99_latency_ms":4015,"time_to_first_token_ms":605,"total_runtime_ms":3402,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0217,"confidence_interval":0.9335,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8525,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3010,"p95_latency_ms":3679,"p99_latency_ms":3946,"time_to_first_token_ms":582,"total_runtime_ms":3344,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0176,"confidence_interval":0.9442,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.8227,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3100,"p95_latency_ms":3789,"p99_latency_ms":4065,"time_to_first_token_ms":586,"total_runtime_ms":3445,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.016,"confidence_interval":0.9386,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.237,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3080,"p95_latency_ms":3765,"p99_latency_ms":4039,"time_to_first_token_ms":579,"total_runtime_ms":3423,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0111,"confidence_interval":0.9407,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7777,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3088,"p95_latency_ms":3774,"p99_latency_ms":4048,"time_to_first_token_ms":580,"total_runtime_ms":3431,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0102,"confidence_interval":0.9497,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.0409,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3078,"p95_latency_ms":3762,"p99_latency_ms":4035,"time_to_first_token_ms":585,"total_runtime_ms":3420,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0125,"confidence_interval":0.9467,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.603,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2997,"p95_latency_ms":3664,"p99_latency_ms":3930,"time_to_first_token_ms":598,"total_runtime_ms":3331,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0269,"confidence_interval":0.9359,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7895,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":3019,"p95_latency_ms":3690,"p99_latency_ms":3958,"time_to_first_token_ms":591,"total_runtime_ms":3354,"cold_start_penalty_ms":1664,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.9484,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2749,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1540,"p95_latency_ms":1882,"p99_latency_ms":2019,"time_to_first_token_ms":550,"total_runtime_ms":1711,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0145,"confidence_interval":0.9389,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5374,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1569,"p95_latency_ms":1917,"p99_latency_ms":2057,"time_to_first_token_ms":548,"total_runtime_ms":1743,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.023,"confidence_interval":0.9401,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.9396,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1604,"p95_latency_ms":1960,"p99_latency_ms":2103,"time_to_first_token_ms":531,"total_runtime_ms":1782,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0104,"confidence_interval":0.9379,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2915,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1513,"p95_latency_ms":1849,"p99_latency_ms":1983,"time_to_first_token_ms":534,"total_runtime_ms":1681,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0156,"confidence_interval":0.9381,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.7665,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1597,"p95_latency_ms":1951,"p99_latency_ms":2093,"time_to_first_token_ms":550,"total_runtime_ms":1774,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0156,"confidence_interval":0.9328,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5109,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1518,"p95_latency_ms":1855,"p99_latency_ms":1990,"time_to_first_token_ms":540,"total_runtime_ms":1687,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0249,"confidence_interval":0.9455,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.4639,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1592,"p95_latency_ms":1945,"p99_latency_ms":2087,"time_to_first_token_ms":555,"total_runtime_ms":1768,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0295,"confidence_interval":0.944,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.5017,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1514,"p95_latency_ms":1850,"p99_latency_ms":1985,"time_to_first_token_ms":539,"total_runtime_ms":1682,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9491,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":11.2818,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1534,"p95_latency_ms":1875,"p99_latency_ms":2012,"time_to_first_token_ms":539,"total_runtime_ms":1705,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9488,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.637,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|low|qwen2:14b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"low","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1552,"p95_latency_ms":1897,"p99_latency_ms":2035,"time_to_first_token_ms":555,"total_runtime_ms":1725,"cold_start_penalty_ms":0,"peak_ram_mb":3584,"avg_ram_mb":2688,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0146,"confidence_interval":0.9421,"validation_overhead_ms":56,"orchestration_overhead_ms":72,"event_dispatch_cost":10.731,"memory_management_cost":12.84,"memory_pressure_events":1} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":592,"p95_latency_ms":723,"p99_latency_ms":776,"time_to_first_token_ms":174,"total_runtime_ms":658,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0264,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7059,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":658,"p95_latency_ms":805,"p99_latency_ms":863,"time_to_first_token_ms":166,"total_runtime_ms":732,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0267,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7358,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":634,"p95_latency_ms":775,"p99_latency_ms":832,"time_to_first_token_ms":168,"total_runtime_ms":705,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0131,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5245,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":646,"p95_latency_ms":789,"p99_latency_ms":847,"time_to_first_token_ms":184,"total_runtime_ms":718,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.025,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0661,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":169,"total_runtime_ms":635,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0112,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3551,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":638,"p95_latency_ms":780,"p99_latency_ms":837,"time_to_first_token_ms":183,"total_runtime_ms":709,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0198,"confidence_interval":0.9454,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4152,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":583,"p95_latency_ms":712,"p99_latency_ms":764,"time_to_first_token_ms":181,"total_runtime_ms":648,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0118,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7238,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":613,"p95_latency_ms":749,"p99_latency_ms":803,"time_to_first_token_ms":183,"total_runtime_ms":681,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0274,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6224,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":622,"p95_latency_ms":760,"p99_latency_ms":815,"time_to_first_token_ms":161,"total_runtime_ms":691,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0288,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8248,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":624,"p95_latency_ms":762,"p99_latency_ms":818,"time_to_first_token_ms":171,"total_runtime_ms":693,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0234,"confidence_interval":0.938,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8952,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":385,"p95_latency_ms":470,"p99_latency_ms":505,"time_to_first_token_ms":138,"total_runtime_ms":428,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0276,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3028,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":349,"p95_latency_ms":426,"p99_latency_ms":457,"time_to_first_token_ms":114,"total_runtime_ms":387,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0139,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4526,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":316,"p95_latency_ms":386,"p99_latency_ms":414,"time_to_first_token_ms":114,"total_runtime_ms":351,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.016,"confidence_interval":0.9336,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.246,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":477,"time_to_first_token_ms":127,"total_runtime_ms":404,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0277,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9446,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":306,"p95_latency_ms":374,"p99_latency_ms":401,"time_to_first_token_ms":142,"total_runtime_ms":340,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0235,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5024,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":314,"p95_latency_ms":384,"p99_latency_ms":412,"time_to_first_token_ms":146,"total_runtime_ms":349,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0144,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6278,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":380,"p95_latency_ms":465,"p99_latency_ms":498,"time_to_first_token_ms":127,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0186,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8141,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":330,"p95_latency_ms":404,"p99_latency_ms":433,"time_to_first_token_ms":124,"total_runtime_ms":367,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0221,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0365,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":374,"p95_latency_ms":457,"p99_latency_ms":491,"time_to_first_token_ms":118,"total_runtime_ms":416,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0209,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4324,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:0.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":328,"p95_latency_ms":401,"p99_latency_ms":431,"time_to_first_token_ms":145,"total_runtime_ms":365,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0155,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0391,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":861,"p95_latency_ms":1053,"p99_latency_ms":1129,"time_to_first_token_ms":211,"total_runtime_ms":957,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0125,"confidence_interval":0.9433,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5186,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":770,"p95_latency_ms":942,"p99_latency_ms":1010,"time_to_first_token_ms":205,"total_runtime_ms":856,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0116,"confidence_interval":0.9492,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6697,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":760,"p95_latency_ms":928,"p99_latency_ms":996,"time_to_first_token_ms":202,"total_runtime_ms":844,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.028,"confidence_interval":0.9426,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8287,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":787,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":203,"total_runtime_ms":874,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0253,"confidence_interval":0.9473,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8659,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":828,"p95_latency_ms":1012,"p99_latency_ms":1086,"time_to_first_token_ms":200,"total_runtime_ms":920,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0217,"confidence_interval":0.9318,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8146,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":863,"p95_latency_ms":1054,"p99_latency_ms":1131,"time_to_first_token_ms":193,"total_runtime_ms":959,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0299,"confidence_interval":0.9306,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7968,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":831,"p95_latency_ms":1016,"p99_latency_ms":1090,"time_to_first_token_ms":211,"total_runtime_ms":924,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0248,"confidence_interval":0.9307,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4074,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":780,"p95_latency_ms":953,"p99_latency_ms":1022,"time_to_first_token_ms":193,"total_runtime_ms":866,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0126,"confidence_interval":0.9496,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5833,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":774,"p95_latency_ms":946,"p99_latency_ms":1015,"time_to_first_token_ms":204,"total_runtime_ms":860,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0225,"confidence_interval":0.9457,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8445,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":806,"p95_latency_ms":985,"p99_latency_ms":1057,"time_to_first_token_ms":197,"total_runtime_ms":896,"cold_start_penalty_ms":421,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.9306,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.6096,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":449,"p95_latency_ms":548,"p99_latency_ms":588,"time_to_first_token_ms":156,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0226,"confidence_interval":0.9486,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8166,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":161,"total_runtime_ms":510,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0127,"confidence_interval":0.9423,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9408,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":471,"p95_latency_ms":576,"p99_latency_ms":618,"time_to_first_token_ms":179,"total_runtime_ms":524,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.015,"confidence_interval":0.9493,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3547,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":455,"p95_latency_ms":556,"p99_latency_ms":596,"time_to_first_token_ms":170,"total_runtime_ms":505,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0266,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2018,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":394,"p95_latency_ms":482,"p99_latency_ms":517,"time_to_first_token_ms":176,"total_runtime_ms":438,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0228,"confidence_interval":0.935,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8904,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":413,"p95_latency_ms":505,"p99_latency_ms":542,"time_to_first_token_ms":173,"total_runtime_ms":459,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9304,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2685,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":484,"p95_latency_ms":592,"p99_latency_ms":635,"time_to_first_token_ms":170,"total_runtime_ms":538,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0247,"confidence_interval":0.9361,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2453,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":447,"p95_latency_ms":547,"p99_latency_ms":587,"time_to_first_token_ms":162,"total_runtime_ms":497,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.9333,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5188,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":166,"total_runtime_ms":504,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0287,"confidence_interval":0.9447,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0647,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:1.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":463,"p95_latency_ms":566,"p99_latency_ms":607,"time_to_first_token_ms":158,"total_runtime_ms":514,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":680,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0252,"confidence_interval":0.9444,"validation_overhead_ms":35,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5366,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1545,"p95_latency_ms":1888,"p99_latency_ms":2026,"time_to_first_token_ms":330,"total_runtime_ms":1717,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0156,"confidence_interval":0.9324,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6967,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1547,"p95_latency_ms":1891,"p99_latency_ms":2029,"time_to_first_token_ms":338,"total_runtime_ms":1719,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.018,"confidence_interval":0.9402,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8227,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1543,"p95_latency_ms":1886,"p99_latency_ms":2023,"time_to_first_token_ms":329,"total_runtime_ms":1714,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0157,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7612,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1587,"p95_latency_ms":1940,"p99_latency_ms":2081,"time_to_first_token_ms":352,"total_runtime_ms":1763,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0215,"confidence_interval":0.9466,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.4509,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1541,"p95_latency_ms":1883,"p99_latency_ms":2020,"time_to_first_token_ms":340,"total_runtime_ms":1712,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.017,"confidence_interval":0.9369,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7294,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1570,"p95_latency_ms":1919,"p99_latency_ms":2058,"time_to_first_token_ms":343,"total_runtime_ms":1744,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0106,"confidence_interval":0.9452,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.7691,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1624,"p95_latency_ms":1984,"p99_latency_ms":2129,"time_to_first_token_ms":341,"total_runtime_ms":1804,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0286,"confidence_interval":0.9448,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.601,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1609,"p95_latency_ms":1966,"p99_latency_ms":2109,"time_to_first_token_ms":328,"total_runtime_ms":1788,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0255,"confidence_interval":0.9353,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1181,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1618,"p95_latency_ms":1978,"p99_latency_ms":2122,"time_to_first_token_ms":350,"total_runtime_ms":1798,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0288,"confidence_interval":0.9305,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5073,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1555,"p95_latency_ms":1901,"p99_latency_ms":2039,"time_to_first_token_ms":350,"total_runtime_ms":1728,"cold_start_penalty_ms":842,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0132,"confidence_interval":0.9416,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.6844,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":840,"p95_latency_ms":1026,"p99_latency_ms":1101,"time_to_first_token_ms":278,"total_runtime_ms":933,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0298,"confidence_interval":0.9392,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1762,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":852,"p95_latency_ms":1042,"p99_latency_ms":1118,"time_to_first_token_ms":285,"total_runtime_ms":947,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0122,"confidence_interval":0.9462,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2259,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":855,"p95_latency_ms":1045,"p99_latency_ms":1121,"time_to_first_token_ms":297,"total_runtime_ms":950,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0222,"confidence_interval":0.9446,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.2854,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":804,"p95_latency_ms":982,"p99_latency_ms":1054,"time_to_first_token_ms":306,"total_runtime_ms":893,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0296,"confidence_interval":0.9358,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9637,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":821,"p95_latency_ms":1004,"p99_latency_ms":1077,"time_to_first_token_ms":302,"total_runtime_ms":913,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0136,"confidence_interval":0.9446,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.5176,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":852,"p95_latency_ms":1041,"p99_latency_ms":1116,"time_to_first_token_ms":298,"total_runtime_ms":946,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.017,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8039,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":834,"p95_latency_ms":1020,"p99_latency_ms":1094,"time_to_first_token_ms":283,"total_runtime_ms":927,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0246,"confidence_interval":0.9314,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.8294,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":836,"p95_latency_ms":1022,"p99_latency_ms":1097,"time_to_first_token_ms":304,"total_runtime_ms":929,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0156,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9283,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":760,"p95_latency_ms":928,"p99_latency_ms":996,"time_to_first_token_ms":285,"total_runtime_ms":844,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0146,"confidence_interval":0.9446,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":8.9345,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:7b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":810,"p95_latency_ms":991,"p99_latency_ms":1063,"time_to_first_token_ms":301,"total_runtime_ms":901,"cold_start_penalty_ms":0,"peak_ram_mb":1814,"avg_ram_mb":1361,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0209,"confidence_interval":0.9327,"validation_overhead_ms":46,"orchestration_overhead_ms":69,"event_dispatch_cost":9.1911,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2263,"p95_latency_ms":2766,"p99_latency_ms":2967,"time_to_first_token_ms":458,"total_runtime_ms":2515,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0238,"confidence_interval":0.9449,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2279,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2354,"p95_latency_ms":2877,"p99_latency_ms":3086,"time_to_first_token_ms":462,"total_runtime_ms":2615,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0223,"confidence_interval":0.9426,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.0193,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2352,"p95_latency_ms":2874,"p99_latency_ms":3083,"time_to_first_token_ms":458,"total_runtime_ms":2613,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0225,"confidence_interval":0.9362,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8139,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2257,"p95_latency_ms":2758,"p99_latency_ms":2959,"time_to_first_token_ms":458,"total_runtime_ms":2507,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0281,"confidence_interval":0.9427,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4529,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2267,"p95_latency_ms":2771,"p99_latency_ms":2973,"time_to_first_token_ms":472,"total_runtime_ms":2519,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0174,"confidence_interval":0.939,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8415,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2335,"p95_latency_ms":2854,"p99_latency_ms":3061,"time_to_first_token_ms":444,"total_runtime_ms":2594,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0298,"confidence_interval":0.9427,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2792,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2250,"p95_latency_ms":2750,"p99_latency_ms":2950,"time_to_first_token_ms":448,"total_runtime_ms":2500,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0286,"confidence_interval":0.938,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8696,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2307,"p95_latency_ms":2820,"p99_latency_ms":3025,"time_to_first_token_ms":473,"total_runtime_ms":2563,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0158,"confidence_interval":0.9326,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.2515,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2293,"p95_latency_ms":2802,"p99_latency_ms":3006,"time_to_first_token_ms":458,"total_runtime_ms":2547,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.9322,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4682,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":2318,"p95_latency_ms":2834,"p99_latency_ms":3040,"time_to_first_token_ms":477,"total_runtime_ms":2576,"cold_start_penalty_ms":1248,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0234,"confidence_interval":0.9409,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.8954,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1125,"p95_latency_ms":1374,"p99_latency_ms":1474,"time_to_first_token_ms":401,"total_runtime_ms":1249,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0234,"confidence_interval":0.935,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.1154,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1139,"p95_latency_ms":1392,"p99_latency_ms":1494,"time_to_first_token_ms":424,"total_runtime_ms":1266,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0101,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.447,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1132,"p95_latency_ms":1383,"p99_latency_ms":1484,"time_to_first_token_ms":410,"total_runtime_ms":1257,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.013,"confidence_interval":0.9486,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7598,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1171,"p95_latency_ms":1431,"p99_latency_ms":1535,"time_to_first_token_ms":424,"total_runtime_ms":1301,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0168,"confidence_interval":0.9479,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.4321,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1213,"p95_latency_ms":1482,"p99_latency_ms":1590,"time_to_first_token_ms":405,"total_runtime_ms":1347,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0229,"confidence_interval":0.9452,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.7605,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1186,"p95_latency_ms":1449,"p99_latency_ms":1555,"time_to_first_token_ms":414,"total_runtime_ms":1318,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0184,"confidence_interval":0.9461,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.118,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1126,"p95_latency_ms":1376,"p99_latency_ms":1476,"time_to_first_token_ms":431,"total_runtime_ms":1251,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.9468,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5186,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1212,"p95_latency_ms":1481,"p99_latency_ms":1589,"time_to_first_token_ms":405,"total_runtime_ms":1347,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0203,"confidence_interval":0.9395,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.5001,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1133,"p95_latency_ms":1384,"p99_latency_ms":1485,"time_to_first_token_ms":406,"total_runtime_ms":1258,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0217,"confidence_interval":0.9459,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":10.775,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium|qwen2:14b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1229,"p95_latency_ms":1503,"p99_latency_ms":1612,"time_to_first_token_ms":403,"total_runtime_ms":1366,"cold_start_penalty_ms":0,"peak_ram_mb":2688,"avg_ram_mb":2016,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0122,"confidence_interval":0.9418,"validation_overhead_ms":56,"orchestration_overhead_ms":69,"event_dispatch_cost":11.007,"memory_management_cost":8.88,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":493,"p95_latency_ms":602,"p99_latency_ms":646,"time_to_first_token_ms":162,"total_runtime_ms":548,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0141,"confidence_interval":0.9346,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9897,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":455,"p95_latency_ms":557,"p99_latency_ms":597,"time_to_first_token_ms":142,"total_runtime_ms":506,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0141,"confidence_interval":0.9416,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5652,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":421,"p95_latency_ms":515,"p99_latency_ms":552,"time_to_first_token_ms":135,"total_runtime_ms":468,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0294,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0244,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":475,"p95_latency_ms":580,"p99_latency_ms":622,"time_to_first_token_ms":138,"total_runtime_ms":527,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0219,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4864,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":423,"p95_latency_ms":518,"p99_latency_ms":555,"time_to_first_token_ms":167,"total_runtime_ms":471,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0193,"confidence_interval":0.9362,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9704,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":516,"p95_latency_ms":631,"p99_latency_ms":676,"time_to_first_token_ms":150,"total_runtime_ms":573,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.021,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1407,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":497,"p95_latency_ms":608,"p99_latency_ms":652,"time_to_first_token_ms":166,"total_runtime_ms":553,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.044,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0277,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1087,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":446,"p95_latency_ms":545,"p99_latency_ms":584,"time_to_first_token_ms":151,"total_runtime_ms":495,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0236,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1201,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":464,"p95_latency_ms":567,"p99_latency_ms":608,"time_to_first_token_ms":166,"total_runtime_ms":516,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.019,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8702,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":480,"p95_latency_ms":586,"p99_latency_ms":629,"time_to_first_token_ms":135,"total_runtime_ms":533,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0157,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3132,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":221,"p95_latency_ms":270,"p99_latency_ms":290,"time_to_first_token_ms":112,"total_runtime_ms":245,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0128,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9075,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":294,"p95_latency_ms":359,"p99_latency_ms":385,"time_to_first_token_ms":111,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0286,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8384,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":229,"p95_latency_ms":280,"p99_latency_ms":301,"time_to_first_token_ms":94,"total_runtime_ms":255,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0236,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3323,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":275,"p95_latency_ms":337,"p99_latency_ms":361,"time_to_first_token_ms":118,"total_runtime_ms":306,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3287,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":318,"p95_latency_ms":389,"p99_latency_ms":417,"time_to_first_token_ms":110,"total_runtime_ms":353,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0127,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0937,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":274,"p95_latency_ms":334,"p99_latency_ms":359,"time_to_first_token_ms":91,"total_runtime_ms":304,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0238,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9197,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":261,"p95_latency_ms":319,"p99_latency_ms":342,"time_to_first_token_ms":96,"total_runtime_ms":290,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0293,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1617,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":260,"p95_latency_ms":317,"p99_latency_ms":341,"time_to_first_token_ms":95,"total_runtime_ms":289,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0283,"confidence_interval":0.9477,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2235,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":267,"p95_latency_ms":326,"p99_latency_ms":350,"time_to_first_token_ms":97,"total_runtime_ms":297,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0156,"confidence_interval":0.95,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8019,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:0.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":238,"p95_latency_ms":290,"p99_latency_ms":311,"time_to_first_token_ms":113,"total_runtime_ms":264,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0289,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0338,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":627,"p95_latency_ms":766,"p99_latency_ms":822,"time_to_first_token_ms":167,"total_runtime_ms":697,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.029,"confidence_interval":0.9315,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7239,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":578,"p95_latency_ms":707,"p99_latency_ms":758,"time_to_first_token_ms":181,"total_runtime_ms":642,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.023,"confidence_interval":0.9481,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7752,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":637,"p95_latency_ms":779,"p99_latency_ms":835,"time_to_first_token_ms":172,"total_runtime_ms":708,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0283,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3842,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":666,"p95_latency_ms":814,"p99_latency_ms":873,"time_to_first_token_ms":176,"total_runtime_ms":740,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.026,"confidence_interval":0.9479,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9152,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":617,"p95_latency_ms":754,"p99_latency_ms":808,"time_to_first_token_ms":174,"total_runtime_ms":685,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0109,"confidence_interval":0.9421,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4005,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":605,"p95_latency_ms":740,"p99_latency_ms":793,"time_to_first_token_ms":177,"total_runtime_ms":672,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0267,"confidence_interval":0.9363,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4788,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":640,"p95_latency_ms":782,"p99_latency_ms":839,"time_to_first_token_ms":172,"total_runtime_ms":711,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0172,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5839,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":636,"p95_latency_ms":777,"p99_latency_ms":834,"time_to_first_token_ms":183,"total_runtime_ms":706,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0264,"confidence_interval":0.943,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0225,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":580,"p95_latency_ms":709,"p99_latency_ms":760,"time_to_first_token_ms":172,"total_runtime_ms":644,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0163,"confidence_interval":0.9352,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.355,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":593,"p95_latency_ms":725,"p99_latency_ms":778,"time_to_first_token_ms":192,"total_runtime_ms":659,"cold_start_penalty_ms":316,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0203,"confidence_interval":0.9347,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8749,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":337,"p95_latency_ms":412,"p99_latency_ms":442,"time_to_first_token_ms":139,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0233,"confidence_interval":0.9388,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.6304,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":380,"p95_latency_ms":464,"p99_latency_ms":498,"time_to_first_token_ms":142,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0288,"confidence_interval":0.9377,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9466,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":331,"p95_latency_ms":405,"p99_latency_ms":434,"time_to_first_token_ms":125,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0269,"confidence_interval":0.9303,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5112,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":347,"p95_latency_ms":424,"p99_latency_ms":455,"time_to_first_token_ms":127,"total_runtime_ms":386,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0299,"confidence_interval":0.9347,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4417,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":384,"p95_latency_ms":469,"p99_latency_ms":503,"time_to_first_token_ms":118,"total_runtime_ms":426,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0222,"confidence_interval":0.9453,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4409,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":315,"p95_latency_ms":385,"p99_latency_ms":413,"time_to_first_token_ms":129,"total_runtime_ms":350,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.024,"confidence_interval":0.9311,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8574,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":356,"p95_latency_ms":435,"p99_latency_ms":466,"time_to_first_token_ms":131,"total_runtime_ms":395,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0137,"confidence_interval":0.9417,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1465,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":290,"p95_latency_ms":355,"p99_latency_ms":381,"time_to_first_token_ms":142,"total_runtime_ms":322,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0121,"confidence_interval":0.9329,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4798,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":337,"p95_latency_ms":412,"p99_latency_ms":442,"time_to_first_token_ms":127,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0226,"confidence_interval":0.9319,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8251,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:1.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":289,"p95_latency_ms":353,"p99_latency_ms":378,"time_to_first_token_ms":131,"total_runtime_ms":321,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":510,"cpu_utilization_percent":29,"gpu_utilization_percent":27,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0236,"confidence_interval":0.9426,"validation_overhead_ms":35,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7852,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1158,"p95_latency_ms":1416,"p99_latency_ms":1519,"time_to_first_token_ms":280,"total_runtime_ms":1287,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0186,"confidence_interval":0.939,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8928,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1205,"p95_latency_ms":1473,"p99_latency_ms":1580,"time_to_first_token_ms":276,"total_runtime_ms":1339,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0191,"confidence_interval":0.9459,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5601,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1157,"p95_latency_ms":1414,"p99_latency_ms":1517,"time_to_first_token_ms":287,"total_runtime_ms":1285,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0247,"confidence_interval":0.9357,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1491,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1210,"p95_latency_ms":1479,"p99_latency_ms":1587,"time_to_first_token_ms":258,"total_runtime_ms":1345,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0268,"confidence_interval":0.9486,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1134,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1151,"p95_latency_ms":1407,"p99_latency_ms":1509,"time_to_first_token_ms":283,"total_runtime_ms":1279,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0181,"confidence_interval":0.9346,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5564,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1168,"p95_latency_ms":1428,"p99_latency_ms":1531,"time_to_first_token_ms":276,"total_runtime_ms":1298,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0206,"confidence_interval":0.9308,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9854,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1140,"p95_latency_ms":1394,"p99_latency_ms":1495,"time_to_first_token_ms":262,"total_runtime_ms":1267,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.01,"confidence_interval":0.9386,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.5335,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1211,"p95_latency_ms":1480,"p99_latency_ms":1588,"time_to_first_token_ms":284,"total_runtime_ms":1346,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0174,"confidence_interval":0.9319,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.284,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1194,"p95_latency_ms":1459,"p99_latency_ms":1566,"time_to_first_token_ms":282,"total_runtime_ms":1327,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0158,"confidence_interval":0.9317,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9542,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1187,"p95_latency_ms":1450,"p99_latency_ms":1556,"time_to_first_token_ms":286,"total_runtime_ms":1318,"cold_start_penalty_ms":632,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":39,"gpu_utilization_percent":54,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.022,"confidence_interval":0.9307,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.854,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":569,"p95_latency_ms":695,"p99_latency_ms":746,"time_to_first_token_ms":213,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0163,"confidence_interval":0.9404,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.0917,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":634,"p95_latency_ms":774,"p99_latency_ms":831,"time_to_first_token_ms":241,"total_runtime_ms":704,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0181,"confidence_interval":0.9405,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6616,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":653,"p95_latency_ms":799,"p99_latency_ms":857,"time_to_first_token_ms":229,"total_runtime_ms":726,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.015,"confidence_interval":0.9449,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1572,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":600,"p95_latency_ms":733,"p99_latency_ms":787,"time_to_first_token_ms":217,"total_runtime_ms":667,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0279,"confidence_interval":0.9401,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.9557,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":569,"p95_latency_ms":696,"p99_latency_ms":746,"time_to_first_token_ms":244,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0269,"confidence_interval":0.9468,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.1444,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":620,"p95_latency_ms":758,"p99_latency_ms":813,"time_to_first_token_ms":220,"total_runtime_ms":689,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0226,"confidence_interval":0.9366,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.8729,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":596,"p95_latency_ms":729,"p99_latency_ms":782,"time_to_first_token_ms":233,"total_runtime_ms":662,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0271,"confidence_interval":0.941,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.349,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":628,"p95_latency_ms":767,"p99_latency_ms":823,"time_to_first_token_ms":216,"total_runtime_ms":698,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.935,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":9.2684,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":756,"time_to_first_token_ms":236,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0241,"confidence_interval":0.9387,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.7117,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:7b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":593,"p95_latency_ms":724,"p99_latency_ms":777,"time_to_first_token_ms":227,"total_runtime_ms":658,"cold_start_penalty_ms":0,"peak_ram_mb":1361,"avg_ram_mb":1021,"cpu_utilization_percent":29,"gpu_utilization_percent":54,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0282,"confidence_interval":0.9336,"validation_overhead_ms":46,"orchestration_overhead_ms":66,"event_dispatch_cost":8.6698,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1716,"p95_latency_ms":2098,"p99_latency_ms":2250,"time_to_first_token_ms":360,"total_runtime_ms":1907,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0156,"confidence_interval":0.9484,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0804,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1696,"p95_latency_ms":2072,"p99_latency_ms":2223,"time_to_first_token_ms":366,"total_runtime_ms":1884,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0208,"confidence_interval":0.9388,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8527,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1763,"p95_latency_ms":2154,"p99_latency_ms":2311,"time_to_first_token_ms":363,"total_runtime_ms":1958,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0261,"confidence_interval":0.9318,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2383,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1762,"p95_latency_ms":2153,"p99_latency_ms":2310,"time_to_first_token_ms":376,"total_runtime_ms":1957,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0249,"confidence_interval":0.9427,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.9302,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1710,"p95_latency_ms":2090,"p99_latency_ms":2242,"time_to_first_token_ms":357,"total_runtime_ms":1900,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0296,"confidence_interval":0.9372,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1604,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1787,"p95_latency_ms":2184,"p99_latency_ms":2343,"time_to_first_token_ms":380,"total_runtime_ms":1985,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0286,"confidence_interval":0.9451,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2665,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1707,"p95_latency_ms":2086,"p99_latency_ms":2238,"time_to_first_token_ms":351,"total_runtime_ms":1897,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0122,"confidence_interval":0.9355,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.303,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1762,"p95_latency_ms":2154,"p99_latency_ms":2310,"time_to_first_token_ms":371,"total_runtime_ms":1958,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.033,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0245,"confidence_interval":0.9306,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.3956,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1697,"p95_latency_ms":2074,"p99_latency_ms":2225,"time_to_first_token_ms":360,"total_runtime_ms":1886,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0212,"confidence_interval":0.9342,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5608,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1722,"p95_latency_ms":2104,"p99_latency_ms":2257,"time_to_first_token_ms":354,"total_runtime_ms":1913,"cold_start_penalty_ms":936,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":39,"gpu_utilization_percent":80,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0162,"confidence_interval":0.9493,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2728,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":872,"p95_latency_ms":1065,"p99_latency_ms":1143,"time_to_first_token_ms":334,"total_runtime_ms":968,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.049,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0221,"confidence_interval":0.9482,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5691,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":943,"p95_latency_ms":1152,"p99_latency_ms":1236,"time_to_first_token_ms":324,"total_runtime_ms":1048,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.032,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0148,"confidence_interval":0.9332,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4691,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":846,"p95_latency_ms":1034,"p99_latency_ms":1109,"time_to_first_token_ms":323,"total_runtime_ms":940,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0102,"confidence_interval":0.9302,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.8857,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":920,"p95_latency_ms":1125,"p99_latency_ms":1206,"time_to_first_token_ms":320,"total_runtime_ms":1022,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0158,"confidence_interval":0.9379,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.5282,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":883,"p95_latency_ms":1080,"p99_latency_ms":1158,"time_to_first_token_ms":304,"total_runtime_ms":982,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0187,"confidence_interval":0.9477,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.1455,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":901,"p95_latency_ms":1101,"p99_latency_ms":1181,"time_to_first_token_ms":330,"total_runtime_ms":1001,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0296,"confidence_interval":0.941,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.0751,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":921,"p95_latency_ms":1125,"p99_latency_ms":1207,"time_to_first_token_ms":315,"total_runtime_ms":1023,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.033,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0224,"confidence_interval":0.9424,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2004,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":901,"p95_latency_ms":1101,"p99_latency_ms":1181,"time_to_first_token_ms":315,"total_runtime_ms":1001,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0125,"confidence_interval":0.931,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":11.2139,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":890,"p95_latency_ms":1087,"p99_latency_ms":1166,"time_to_first_token_ms":323,"total_runtime_ms":988,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0264,"confidence_interval":0.9405,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.4102,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|medium_gpu|qwen2:14b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":899,"p95_latency_ms":1099,"p99_latency_ms":1179,"time_to_first_token_ms":309,"total_runtime_ms":999,"cold_start_penalty_ms":0,"peak_ram_mb":2016,"avg_ram_mb":1512,"cpu_utilization_percent":29,"gpu_utilization_percent":80,"oscillation_rate":0.044,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0112,"confidence_interval":0.9486,"validation_overhead_ms":56,"orchestration_overhead_ms":66,"event_dispatch_cost":10.7742,"memory_management_cost":8.16,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":449,"p95_latency_ms":549,"p99_latency_ms":589,"time_to_first_token_ms":133,"total_runtime_ms":499,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0236,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.376,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":353,"p95_latency_ms":432,"p99_latency_ms":463,"time_to_first_token_ms":124,"total_runtime_ms":393,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0283,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8473,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":555,"time_to_first_token_ms":130,"total_runtime_ms":470,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0177,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9504,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":394,"p95_latency_ms":482,"p99_latency_ms":517,"time_to_first_token_ms":150,"total_runtime_ms":438,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0185,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4674,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":364,"p95_latency_ms":444,"p99_latency_ms":477,"time_to_first_token_ms":139,"total_runtime_ms":404,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0142,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.674,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":452,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":128,"total_runtime_ms":503,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0217,"confidence_interval":0.9305,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1436,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":442,"p95_latency_ms":540,"p99_latency_ms":579,"time_to_first_token_ms":133,"total_runtime_ms":491,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0255,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2776,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":385,"p95_latency_ms":471,"p99_latency_ms":505,"time_to_first_token_ms":123,"total_runtime_ms":428,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0246,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2995,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":407,"p95_latency_ms":498,"p99_latency_ms":534,"time_to_first_token_ms":144,"total_runtime_ms":453,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0212,"confidence_interval":0.9336,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7451,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":363,"p95_latency_ms":444,"p99_latency_ms":476,"time_to_first_token_ms":132,"total_runtime_ms":404,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0271,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6165,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":279,"p95_latency_ms":340,"p99_latency_ms":365,"time_to_first_token_ms":89,"total_runtime_ms":310,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0194,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5784,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":264,"p95_latency_ms":322,"p99_latency_ms":346,"time_to_first_token_ms":98,"total_runtime_ms":293,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0143,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8682,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":278,"p95_latency_ms":340,"p99_latency_ms":364,"time_to_first_token_ms":107,"total_runtime_ms":309,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0259,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3664,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":185,"p95_latency_ms":227,"p99_latency_ms":243,"time_to_first_token_ms":95,"total_runtime_ms":206,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0152,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4092,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":215,"p95_latency_ms":263,"p99_latency_ms":282,"time_to_first_token_ms":88,"total_runtime_ms":239,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.02,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5646,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":224,"p95_latency_ms":274,"p99_latency_ms":294,"time_to_first_token_ms":90,"total_runtime_ms":249,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0176,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8128,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":262,"p95_latency_ms":320,"p99_latency_ms":343,"time_to_first_token_ms":77,"total_runtime_ms":291,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0273,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0502,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":268,"p95_latency_ms":327,"p99_latency_ms":351,"time_to_first_token_ms":96,"total_runtime_ms":297,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0163,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9094,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":257,"p95_latency_ms":314,"p99_latency_ms":337,"time_to_first_token_ms":84,"total_runtime_ms":286,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0253,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8966,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:0.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:0.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":262,"p95_latency_ms":320,"p99_latency_ms":343,"time_to_first_token_ms":84,"total_runtime_ms":291,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.018,"confidence_interval":0.9452,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7297,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":495,"p95_latency_ms":606,"p99_latency_ms":650,"time_to_first_token_ms":151,"total_runtime_ms":550,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0149,"confidence_interval":0.9495,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.631,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":558,"p95_latency_ms":682,"p99_latency_ms":732,"time_to_first_token_ms":156,"total_runtime_ms":620,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0195,"confidence_interval":0.9448,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4241,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":578,"p95_latency_ms":706,"p99_latency_ms":758,"time_to_first_token_ms":153,"total_runtime_ms":642,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0114,"confidence_interval":0.9495,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.102,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":548,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":145,"total_runtime_ms":609,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0109,"confidence_interval":0.9312,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0404,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":479,"p95_latency_ms":585,"p99_latency_ms":628,"time_to_first_token_ms":170,"total_runtime_ms":532,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0105,"confidence_interval":0.9376,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9904,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":672,"time_to_first_token_ms":145,"total_runtime_ms":569,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0281,"confidence_interval":0.9365,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5476,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":477,"p95_latency_ms":582,"p99_latency_ms":625,"time_to_first_token_ms":172,"total_runtime_ms":529,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0107,"confidence_interval":0.9328,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9585,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":147,"total_runtime_ms":576,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0182,"confidence_interval":0.9447,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5704,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":509,"p95_latency_ms":622,"p99_latency_ms":667,"time_to_first_token_ms":164,"total_runtime_ms":565,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0279,"confidence_interval":0.9415,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8231,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":553,"p95_latency_ms":676,"p99_latency_ms":726,"time_to_first_token_ms":159,"total_runtime_ms":615,"cold_start_penalty_ms":263,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.022,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.733,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":408,"time_to_first_token_ms":130,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0121,"confidence_interval":0.9349,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.6003,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":242,"p95_latency_ms":296,"p99_latency_ms":317,"time_to_first_token_ms":110,"total_runtime_ms":269,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.014,"confidence_interval":0.9485,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3626,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":313,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":117,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0128,"confidence_interval":0.9407,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3631,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":320,"p95_latency_ms":391,"p99_latency_ms":420,"time_to_first_token_ms":126,"total_runtime_ms":356,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.019,"confidence_interval":0.9351,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4056,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":258,"p95_latency_ms":316,"p99_latency_ms":339,"time_to_first_token_ms":128,"total_runtime_ms":287,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.014,"confidence_interval":0.94,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9526,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":309,"p95_latency_ms":378,"p99_latency_ms":405,"time_to_first_token_ms":117,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0239,"confidence_interval":0.9308,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1106,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":279,"p95_latency_ms":341,"p99_latency_ms":366,"time_to_first_token_ms":108,"total_runtime_ms":310,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0289,"confidence_interval":0.9449,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5174,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":314,"p95_latency_ms":383,"p99_latency_ms":411,"time_to_first_token_ms":131,"total_runtime_ms":349,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0154,"confidence_interval":0.9483,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5279,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":281,"p95_latency_ms":343,"p99_latency_ms":368,"time_to_first_token_ms":98,"total_runtime_ms":312,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0165,"confidence_interval":0.9345,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5191,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:1.5b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:1.5b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":242,"p95_latency_ms":296,"p99_latency_ms":317,"time_to_first_token_ms":111,"total_runtime_ms":269,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":425,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.9,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0252,"confidence_interval":0.9387,"validation_overhead_ms":35,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4633,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1036,"p95_latency_ms":1267,"p99_latency_ms":1359,"time_to_first_token_ms":226,"total_runtime_ms":1152,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0135,"confidence_interval":0.9412,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2117,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":973,"p95_latency_ms":1189,"p99_latency_ms":1276,"time_to_first_token_ms":247,"total_runtime_ms":1081,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0106,"confidence_interval":0.9423,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9924,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":955,"p95_latency_ms":1167,"p99_latency_ms":1252,"time_to_first_token_ms":242,"total_runtime_ms":1061,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0224,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.058,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1013,"p95_latency_ms":1238,"p99_latency_ms":1328,"time_to_first_token_ms":229,"total_runtime_ms":1125,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0177,"confidence_interval":0.9469,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1755,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":992,"p95_latency_ms":1213,"p99_latency_ms":1301,"time_to_first_token_ms":246,"total_runtime_ms":1102,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0279,"confidence_interval":0.9427,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7458,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1008,"p95_latency_ms":1232,"p99_latency_ms":1322,"time_to_first_token_ms":222,"total_runtime_ms":1120,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.027,"confidence_interval":0.9436,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.715,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1048,"p95_latency_ms":1281,"p99_latency_ms":1374,"time_to_first_token_ms":255,"total_runtime_ms":1165,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.013,"confidence_interval":0.9334,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1687,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":971,"p95_latency_ms":1187,"p99_latency_ms":1273,"time_to_first_token_ms":238,"total_runtime_ms":1079,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0123,"confidence_interval":0.9494,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.7303,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":988,"p95_latency_ms":1208,"p99_latency_ms":1295,"time_to_first_token_ms":245,"total_runtime_ms":1098,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0211,"confidence_interval":0.9476,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3817,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":970,"p95_latency_ms":1186,"p99_latency_ms":1272,"time_to_first_token_ms":253,"total_runtime_ms":1078,"cold_start_penalty_ms":527,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0231,"confidence_interval":0.9498,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.9326,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":552,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":201,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.9481,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.1216,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":497,"p95_latency_ms":608,"p99_latency_ms":652,"time_to_first_token_ms":180,"total_runtime_ms":552,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0142,"confidence_interval":0.9363,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3603,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":559,"p95_latency_ms":683,"p99_latency_ms":732,"time_to_first_token_ms":198,"total_runtime_ms":621,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0201,"confidence_interval":0.9345,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.2174,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":548,"p95_latency_ms":670,"p99_latency_ms":719,"time_to_first_token_ms":200,"total_runtime_ms":609,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0286,"confidence_interval":0.9354,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.4251,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":668,"time_to_first_token_ms":185,"total_runtime_ms":567,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0192,"confidence_interval":0.9333,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.214,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":492,"p95_latency_ms":601,"p99_latency_ms":645,"time_to_first_token_ms":191,"total_runtime_ms":546,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0196,"confidence_interval":0.9453,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":9.0887,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":520,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":179,"total_runtime_ms":578,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0146,"confidence_interval":0.9346,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.922,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":756,"time_to_first_token_ms":179,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0218,"confidence_interval":0.9413,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.6937,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":552,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":179,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0289,"confidence_interval":0.9435,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.8388,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:7b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:7b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":529,"p95_latency_ms":646,"p99_latency_ms":693,"time_to_first_token_ms":181,"total_runtime_ms":588,"cold_start_penalty_ms":0,"peak_ram_mb":1134,"avg_ram_mb":851,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.16,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0171,"confidence_interval":0.9406,"validation_overhead_ms":46,"orchestration_overhead_ms":65,"event_dispatch_cost":8.3849,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1459,"p95_latency_ms":1784,"p99_latency_ms":1913,"time_to_first_token_ms":333,"total_runtime_ms":1622,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0277,"confidence_interval":0.9413,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.9252,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1483,"p95_latency_ms":1812,"p99_latency_ms":1944,"time_to_first_token_ms":315,"total_runtime_ms":1647,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0136,"confidence_interval":0.9327,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0078,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1498,"p95_latency_ms":1831,"p99_latency_ms":1965,"time_to_first_token_ms":324,"total_runtime_ms":1665,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0136,"confidence_interval":0.9341,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5622,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1433,"p95_latency_ms":1752,"p99_latency_ms":1879,"time_to_first_token_ms":308,"total_runtime_ms":1593,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0186,"confidence_interval":0.9462,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4482,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1473,"p95_latency_ms":1800,"p99_latency_ms":1931,"time_to_first_token_ms":307,"total_runtime_ms":1637,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0107,"confidence_interval":0.935,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6503,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1418,"p95_latency_ms":1733,"p99_latency_ms":1859,"time_to_first_token_ms":319,"total_runtime_ms":1576,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0248,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4224,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1471,"p95_latency_ms":1798,"p99_latency_ms":1929,"time_to_first_token_ms":313,"total_runtime_ms":1634,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0259,"confidence_interval":0.942,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.2989,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1435,"p95_latency_ms":1754,"p99_latency_ms":1881,"time_to_first_token_ms":316,"total_runtime_ms":1594,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0199,"confidence_interval":0.9396,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5181,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1437,"p95_latency_ms":1756,"p99_latency_ms":1883,"time_to_first_token_ms":321,"total_runtime_ms":1596,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.9481,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.8389,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.85,"schema_validity_rate":0.98,"parse_success_rate":0.83,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.15,"validation_failure_rate":0.075,"p50_latency_ms":1484,"p95_latency_ms":1814,"p99_latency_ms":1946,"time_to_first_token_ms":307,"total_runtime_ms":1649,"cold_start_penalty_ms":780,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0225,"confidence_interval":0.9346,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.7569,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":792,"p95_latency_ms":968,"p99_latency_ms":1038,"time_to_first_token_ms":255,"total_runtime_ms":880,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0178,"confidence_interval":0.9312,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0657,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":808,"p95_latency_ms":988,"p99_latency_ms":1059,"time_to_first_token_ms":289,"total_runtime_ms":898,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0161,"confidence_interval":0.9362,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.1244,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":783,"p95_latency_ms":957,"p99_latency_ms":1027,"time_to_first_token_ms":255,"total_runtime_ms":870,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0224,"confidence_interval":0.9404,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.0533,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":277,"total_runtime_ms":889,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.013,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.6978,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":752,"p95_latency_ms":919,"p99_latency_ms":986,"time_to_first_token_ms":257,"total_runtime_ms":836,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0106,"confidence_interval":0.9302,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.001,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":727,"p95_latency_ms":889,"p99_latency_ms":953,"time_to_first_token_ms":257,"total_runtime_ms":808,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0239,"confidence_interval":0.9367,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3952,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":734,"p95_latency_ms":897,"p99_latency_ms":962,"time_to_first_token_ms":263,"total_runtime_ms":815,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.017,"confidence_interval":0.9469,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.4157,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":760,"p95_latency_ms":929,"p99_latency_ms":997,"time_to_first_token_ms":288,"total_runtime_ms":845,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0243,"confidence_interval":0.9419,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.5353,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":778,"p95_latency_ms":951,"p99_latency_ms":1020,"time_to_first_token_ms":257,"total_runtime_ms":864,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0281,"confidence_interval":0.9311,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":10.583,"memory_management_cost":7.8,"memory_pressure_events":0} -{"combo_id":"noise_input|high|qwen2:14b|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"high","model":"qwen2:14b","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":764,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":263,"total_runtime_ms":848,"cold_start_penalty_ms":0,"peak_ram_mb":1680,"avg_ram_mb":1260,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.3,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0248,"confidence_interval":0.9497,"validation_overhead_ms":56,"orchestration_overhead_ms":65,"event_dispatch_cost":11.3448,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":189,"total_runtime_ms":889,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0199,"confidence_interval":0.9454,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1818,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":752,"p95_latency_ms":919,"p99_latency_ms":986,"time_to_first_token_ms":217,"total_runtime_ms":835,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0295,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5978,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":763,"p95_latency_ms":932,"p99_latency_ms":1000,"time_to_first_token_ms":189,"total_runtime_ms":847,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.024,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2984,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":855,"p95_latency_ms":1045,"p99_latency_ms":1121,"time_to_first_token_ms":200,"total_runtime_ms":950,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0282,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8935,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":799,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":192,"total_runtime_ms":887,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0189,"confidence_interval":0.9395,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4228,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":205,"total_runtime_ms":895,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0109,"confidence_interval":0.9408,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3701,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":793,"p95_latency_ms":969,"p99_latency_ms":1040,"time_to_first_token_ms":219,"total_runtime_ms":881,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0246,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3493,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":771,"p95_latency_ms":942,"p99_latency_ms":1010,"time_to_first_token_ms":223,"total_runtime_ms":856,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0162,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4093,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":832,"p95_latency_ms":1017,"p99_latency_ms":1091,"time_to_first_token_ms":206,"total_runtime_ms":924,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0109,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4748,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":819,"p95_latency_ms":1001,"p99_latency_ms":1074,"time_to_first_token_ms":195,"total_runtime_ms":910,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0139,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.144,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":144,"total_runtime_ms":510,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0139,"confidence_interval":0.9479,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4405,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":152,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0248,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7156,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":458,"p95_latency_ms":560,"p99_latency_ms":601,"time_to_first_token_ms":160,"total_runtime_ms":509,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0251,"confidence_interval":0.9319,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3014,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":380,"p95_latency_ms":465,"p99_latency_ms":498,"time_to_first_token_ms":165,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0284,"confidence_interval":0.9346,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2123,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":405,"p95_latency_ms":494,"p99_latency_ms":530,"time_to_first_token_ms":149,"total_runtime_ms":450,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0794,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":394,"p95_latency_ms":482,"p99_latency_ms":517,"time_to_first_token_ms":148,"total_runtime_ms":438,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.025,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8802,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":168,"total_runtime_ms":504,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0106,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0797,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":379,"p95_latency_ms":463,"p99_latency_ms":496,"time_to_first_token_ms":148,"total_runtime_ms":421,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0258,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0023,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":453,"p95_latency_ms":554,"p99_latency_ms":594,"time_to_first_token_ms":167,"total_runtime_ms":503,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0292,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2126,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":472,"p95_latency_ms":577,"p99_latency_ms":619,"time_to_first_token_ms":157,"total_runtime_ms":525,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0105,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.194,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1304,"p95_latency_ms":1594,"p99_latency_ms":1710,"time_to_first_token_ms":306,"total_runtime_ms":1449,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0266,"confidence_interval":0.9418,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3044,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1304,"p95_latency_ms":1593,"p99_latency_ms":1709,"time_to_first_token_ms":304,"total_runtime_ms":1448,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.941,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0694,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1377,"p95_latency_ms":1683,"p99_latency_ms":1806,"time_to_first_token_ms":300,"total_runtime_ms":1530,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0294,"confidence_interval":0.945,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7542,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1349,"p95_latency_ms":1648,"p99_latency_ms":1768,"time_to_first_token_ms":280,"total_runtime_ms":1499,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0226,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4698,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1324,"p95_latency_ms":1619,"p99_latency_ms":1736,"time_to_first_token_ms":309,"total_runtime_ms":1472,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0241,"confidence_interval":0.9324,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3732,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1275,"p95_latency_ms":1558,"p99_latency_ms":1671,"time_to_first_token_ms":295,"total_runtime_ms":1416,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0192,"confidence_interval":0.9414,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7211,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1277,"p95_latency_ms":1560,"p99_latency_ms":1674,"time_to_first_token_ms":308,"total_runtime_ms":1419,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0169,"confidence_interval":0.9366,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4291,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1324,"p95_latency_ms":1618,"p99_latency_ms":1735,"time_to_first_token_ms":311,"total_runtime_ms":1471,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0287,"confidence_interval":0.9327,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6094,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1348,"p95_latency_ms":1648,"p99_latency_ms":1768,"time_to_first_token_ms":307,"total_runtime_ms":1498,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0164,"confidence_interval":0.9426,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0443,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1329,"p95_latency_ms":1624,"p99_latency_ms":1742,"time_to_first_token_ms":311,"total_runtime_ms":1477,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0115,"confidence_interval":0.9335,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5234,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":717,"p95_latency_ms":876,"p99_latency_ms":940,"time_to_first_token_ms":246,"total_runtime_ms":796,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.012,"confidence_interval":0.9364,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5222,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":643,"p95_latency_ms":785,"p99_latency_ms":842,"time_to_first_token_ms":233,"total_runtime_ms":714,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0178,"confidence_interval":0.9346,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1054,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":681,"p95_latency_ms":832,"p99_latency_ms":892,"time_to_first_token_ms":234,"total_runtime_ms":756,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0108,"confidence_interval":0.9434,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5004,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":646,"p95_latency_ms":789,"p99_latency_ms":847,"time_to_first_token_ms":245,"total_runtime_ms":718,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0291,"confidence_interval":0.944,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9647,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":662,"p95_latency_ms":810,"p99_latency_ms":868,"time_to_first_token_ms":243,"total_runtime_ms":736,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.018,"confidence_interval":0.9483,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6027,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":715,"p95_latency_ms":874,"p99_latency_ms":937,"time_to_first_token_ms":263,"total_runtime_ms":794,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0268,"confidence_interval":0.9405,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5308,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":644,"p95_latency_ms":787,"p99_latency_ms":844,"time_to_first_token_ms":256,"total_runtime_ms":715,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.025,"confidence_interval":0.9419,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6912,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":681,"p95_latency_ms":833,"p99_latency_ms":893,"time_to_first_token_ms":264,"total_runtime_ms":757,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0183,"confidence_interval":0.9426,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4599,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":660,"p95_latency_ms":807,"p99_latency_ms":865,"time_to_first_token_ms":253,"total_runtime_ms":733,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0217,"confidence_interval":0.9367,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2233,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":648,"p95_latency_ms":792,"p99_latency_ms":850,"time_to_first_token_ms":252,"total_runtime_ms":720,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0249,"confidence_interval":0.9404,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5096,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":567,"p95_latency_ms":694,"p99_latency_ms":744,"time_to_first_token_ms":163,"total_runtime_ms":631,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0243,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6755,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":657,"p95_latency_ms":803,"p99_latency_ms":862,"time_to_first_token_ms":177,"total_runtime_ms":730,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0284,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8601,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":610,"p95_latency_ms":745,"p99_latency_ms":799,"time_to_first_token_ms":185,"total_runtime_ms":678,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0144,"confidence_interval":0.9438,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6585,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":578,"p95_latency_ms":706,"p99_latency_ms":757,"time_to_first_token_ms":179,"total_runtime_ms":642,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0165,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9975,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":631,"p95_latency_ms":771,"p99_latency_ms":827,"time_to_first_token_ms":161,"total_runtime_ms":701,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0248,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6379,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":652,"p95_latency_ms":796,"p99_latency_ms":854,"time_to_first_token_ms":168,"total_runtime_ms":724,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0178,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9794,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":634,"p95_latency_ms":775,"p99_latency_ms":832,"time_to_first_token_ms":184,"total_runtime_ms":705,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0196,"confidence_interval":0.9307,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9217,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":615,"p95_latency_ms":751,"p99_latency_ms":806,"time_to_first_token_ms":157,"total_runtime_ms":683,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0215,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4054,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":593,"p95_latency_ms":725,"p99_latency_ms":778,"time_to_first_token_ms":165,"total_runtime_ms":659,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0145,"confidence_interval":0.93,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4985,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":785,"time_to_first_token_ms":182,"total_runtime_ms":665,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2589,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":406,"time_to_first_token_ms":125,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0162,"confidence_interval":0.9465,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8712,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":407,"time_to_first_token_ms":127,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0146,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1169,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":291,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":116,"total_runtime_ms":323,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0219,"confidence_interval":0.9378,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6748,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":363,"p95_latency_ms":443,"p99_latency_ms":475,"time_to_first_token_ms":125,"total_runtime_ms":403,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.023,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.254,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":281,"p95_latency_ms":344,"p99_latency_ms":369,"time_to_first_token_ms":123,"total_runtime_ms":312,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0106,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0224,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":287,"p95_latency_ms":350,"p99_latency_ms":376,"time_to_first_token_ms":118,"total_runtime_ms":318,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.012,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8305,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":306,"p95_latency_ms":374,"p99_latency_ms":401,"time_to_first_token_ms":120,"total_runtime_ms":340,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0105,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7438,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":338,"p95_latency_ms":413,"p99_latency_ms":443,"time_to_first_token_ms":128,"total_runtime_ms":376,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0102,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7411,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":356,"p95_latency_ms":435,"p99_latency_ms":467,"time_to_first_token_ms":114,"total_runtime_ms":396,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0226,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9739,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":322,"p95_latency_ms":393,"p99_latency_ms":422,"time_to_first_token_ms":134,"total_runtime_ms":357,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0217,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3473,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":970,"p95_latency_ms":1185,"p99_latency_ms":1271,"time_to_first_token_ms":233,"total_runtime_ms":1077,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0139,"confidence_interval":0.9381,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8852,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":971,"p95_latency_ms":1187,"p99_latency_ms":1273,"time_to_first_token_ms":238,"total_runtime_ms":1079,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0137,"confidence_interval":0.9326,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.539,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1049,"p95_latency_ms":1283,"p99_latency_ms":1376,"time_to_first_token_ms":254,"total_runtime_ms":1166,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0272,"confidence_interval":0.9407,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2753,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":984,"p95_latency_ms":1202,"p99_latency_ms":1290,"time_to_first_token_ms":253,"total_runtime_ms":1093,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0277,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8119,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":978,"p95_latency_ms":1195,"p99_latency_ms":1282,"time_to_first_token_ms":253,"total_runtime_ms":1086,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0119,"confidence_interval":0.9385,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0046,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":972,"p95_latency_ms":1187,"p99_latency_ms":1274,"time_to_first_token_ms":243,"total_runtime_ms":1079,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0284,"confidence_interval":0.9451,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6315,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1047,"p95_latency_ms":1279,"p99_latency_ms":1372,"time_to_first_token_ms":228,"total_runtime_ms":1163,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0211,"confidence_interval":0.9321,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4347,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1035,"p95_latency_ms":1265,"p99_latency_ms":1357,"time_to_first_token_ms":241,"total_runtime_ms":1150,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.029,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.972,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":958,"p95_latency_ms":1171,"p99_latency_ms":1256,"time_to_first_token_ms":250,"total_runtime_ms":1065,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0132,"confidence_interval":0.9356,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5273,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1005,"p95_latency_ms":1228,"p99_latency_ms":1318,"time_to_first_token_ms":223,"total_runtime_ms":1117,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0223,"confidence_interval":0.9333,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2997,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":486,"p95_latency_ms":594,"p99_latency_ms":637,"time_to_first_token_ms":184,"total_runtime_ms":540,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0189,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2455,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":506,"p95_latency_ms":618,"p99_latency_ms":663,"time_to_first_token_ms":206,"total_runtime_ms":562,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0161,"confidence_interval":0.9322,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7381,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":498,"p95_latency_ms":608,"p99_latency_ms":653,"time_to_first_token_ms":196,"total_runtime_ms":553,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0131,"confidence_interval":0.9399,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4428,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":502,"p95_latency_ms":613,"p99_latency_ms":658,"time_to_first_token_ms":187,"total_runtime_ms":557,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0124,"confidence_interval":0.949,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3577,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":574,"p95_latency_ms":701,"p99_latency_ms":753,"time_to_first_token_ms":212,"total_runtime_ms":638,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0214,"confidence_interval":0.9381,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2713,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":537,"p95_latency_ms":657,"p99_latency_ms":704,"time_to_first_token_ms":187,"total_runtime_ms":597,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0295,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.083,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":744,"time_to_first_token_ms":212,"total_runtime_ms":630,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0193,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.829,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":523,"p95_latency_ms":639,"p99_latency_ms":686,"time_to_first_token_ms":192,"total_runtime_ms":581,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.016,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2089,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":525,"p95_latency_ms":642,"p99_latency_ms":688,"time_to_first_token_ms":207,"total_runtime_ms":583,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0141,"confidence_interval":0.9395,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4661,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":198,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0266,"confidence_interval":0.9331,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6589,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":444,"p95_latency_ms":543,"p99_latency_ms":582,"time_to_first_token_ms":151,"total_runtime_ms":493,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0251,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6954,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":450,"p95_latency_ms":549,"p99_latency_ms":589,"time_to_first_token_ms":133,"total_runtime_ms":500,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4499,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":453,"p95_latency_ms":554,"p99_latency_ms":594,"time_to_first_token_ms":147,"total_runtime_ms":504,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0151,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9774,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":150,"total_runtime_ms":584,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0209,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4431,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":498,"p95_latency_ms":609,"p99_latency_ms":653,"time_to_first_token_ms":161,"total_runtime_ms":553,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0162,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4691,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":437,"p95_latency_ms":534,"p99_latency_ms":572,"time_to_first_token_ms":148,"total_runtime_ms":485,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0211,"confidence_interval":0.942,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0974,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":446,"p95_latency_ms":545,"p99_latency_ms":585,"time_to_first_token_ms":144,"total_runtime_ms":495,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0113,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2093,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":472,"p95_latency_ms":577,"p99_latency_ms":619,"time_to_first_token_ms":155,"total_runtime_ms":524,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0222,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6001,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":505,"p95_latency_ms":617,"p99_latency_ms":661,"time_to_first_token_ms":164,"total_runtime_ms":561,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0205,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1321,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":447,"p95_latency_ms":547,"p99_latency_ms":586,"time_to_first_token_ms":156,"total_runtime_ms":497,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.022,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3069,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":223,"p95_latency_ms":272,"p99_latency_ms":292,"time_to_first_token_ms":98,"total_runtime_ms":248,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0203,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.028,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":303,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":100,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0131,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5052,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":272,"p95_latency_ms":333,"p99_latency_ms":357,"time_to_first_token_ms":87,"total_runtime_ms":302,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0119,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3847,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":277,"p95_latency_ms":338,"p99_latency_ms":363,"time_to_first_token_ms":102,"total_runtime_ms":307,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0111,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.051,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":222,"p95_latency_ms":271,"p99_latency_ms":291,"time_to_first_token_ms":109,"total_runtime_ms":247,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0125,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7622,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":239,"p95_latency_ms":292,"p99_latency_ms":313,"time_to_first_token_ms":109,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0227,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8738,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":283,"p95_latency_ms":346,"p99_latency_ms":371,"time_to_first_token_ms":109,"total_runtime_ms":315,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0116,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0477,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":297,"p95_latency_ms":363,"p99_latency_ms":389,"time_to_first_token_ms":114,"total_runtime_ms":330,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0126,"confidence_interval":0.9424,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6332,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":284,"p95_latency_ms":347,"p99_latency_ms":372,"time_to_first_token_ms":90,"total_runtime_ms":315,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0297,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4607,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":251,"p95_latency_ms":307,"p99_latency_ms":329,"time_to_first_token_ms":91,"total_runtime_ms":279,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0262,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1013,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":806,"p95_latency_ms":985,"p99_latency_ms":1056,"time_to_first_token_ms":191,"total_runtime_ms":895,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0237,"confidence_interval":0.9467,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3742,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":778,"p95_latency_ms":951,"p99_latency_ms":1020,"time_to_first_token_ms":208,"total_runtime_ms":864,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0173,"confidence_interval":0.9462,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6694,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":814,"p95_latency_ms":995,"p99_latency_ms":1067,"time_to_first_token_ms":208,"total_runtime_ms":905,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0181,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0194,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":812,"p95_latency_ms":993,"p99_latency_ms":1065,"time_to_first_token_ms":186,"total_runtime_ms":903,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0146,"confidence_interval":0.9365,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8486,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":743,"p95_latency_ms":908,"p99_latency_ms":974,"time_to_first_token_ms":197,"total_runtime_ms":825,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0139,"confidence_interval":0.9404,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8176,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":948,"time_to_first_token_ms":191,"total_runtime_ms":803,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.018,"confidence_interval":0.947,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1223,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":761,"p95_latency_ms":930,"p99_latency_ms":997,"time_to_first_token_ms":190,"total_runtime_ms":845,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0147,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4438,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":748,"p95_latency_ms":914,"p99_latency_ms":980,"time_to_first_token_ms":208,"total_runtime_ms":831,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.03,"confidence_interval":0.9385,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6061,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":744,"p95_latency_ms":910,"p99_latency_ms":976,"time_to_first_token_ms":215,"total_runtime_ms":827,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0202,"confidence_interval":0.9336,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3134,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":725,"p95_latency_ms":887,"p99_latency_ms":951,"time_to_first_token_ms":184,"total_runtime_ms":806,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0229,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.682,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":548,"time_to_first_token_ms":144,"total_runtime_ms":464,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0235,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9167,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":438,"p95_latency_ms":535,"p99_latency_ms":574,"time_to_first_token_ms":155,"total_runtime_ms":487,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0283,"confidence_interval":0.9441,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4953,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":379,"p95_latency_ms":463,"p99_latency_ms":497,"time_to_first_token_ms":160,"total_runtime_ms":421,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0211,"confidence_interval":0.9396,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2195,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":435,"p95_latency_ms":532,"p99_latency_ms":571,"time_to_first_token_ms":165,"total_runtime_ms":484,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0155,"confidence_interval":0.9347,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3459,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":461,"p95_latency_ms":563,"p99_latency_ms":604,"time_to_first_token_ms":152,"total_runtime_ms":512,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.013,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6046,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":452,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":168,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0224,"confidence_interval":0.9367,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1897,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":162,"total_runtime_ms":453,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0132,"confidence_interval":0.9466,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8078,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":447,"p95_latency_ms":547,"p99_latency_ms":586,"time_to_first_token_ms":152,"total_runtime_ms":497,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0194,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9648,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":363,"p95_latency_ms":444,"p99_latency_ms":476,"time_to_first_token_ms":167,"total_runtime_ms":404,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0228,"confidence_interval":0.94,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.965,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":404,"p95_latency_ms":493,"p99_latency_ms":529,"time_to_first_token_ms":149,"total_runtime_ms":449,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0208,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5668,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":416,"p95_latency_ms":509,"p99_latency_ms":546,"time_to_first_token_ms":128,"total_runtime_ms":463,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0211,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3767,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":425,"p95_latency_ms":519,"p99_latency_ms":557,"time_to_first_token_ms":150,"total_runtime_ms":472,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0271,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4904,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":471,"time_to_first_token_ms":154,"total_runtime_ms":399,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0204,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4907,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":373,"p95_latency_ms":456,"p99_latency_ms":489,"time_to_first_token_ms":131,"total_runtime_ms":415,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.03,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9447,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":143,"total_runtime_ms":504,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0184,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0184,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":398,"p95_latency_ms":486,"p99_latency_ms":521,"time_to_first_token_ms":140,"total_runtime_ms":442,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0298,"confidence_interval":0.9362,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5023,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":379,"p95_latency_ms":463,"p99_latency_ms":496,"time_to_first_token_ms":146,"total_runtime_ms":421,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0114,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5395,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":420,"p95_latency_ms":514,"p99_latency_ms":551,"time_to_first_token_ms":138,"total_runtime_ms":467,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0257,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4928,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":555,"time_to_first_token_ms":129,"total_runtime_ms":470,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0287,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4985,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":452,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":149,"total_runtime_ms":502,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0254,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8873,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":265,"p95_latency_ms":324,"p99_latency_ms":347,"time_to_first_token_ms":100,"total_runtime_ms":294,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0277,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7071,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":195,"p95_latency_ms":238,"p99_latency_ms":255,"time_to_first_token_ms":80,"total_runtime_ms":216,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0106,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.765,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":180,"p95_latency_ms":220,"p99_latency_ms":236,"time_to_first_token_ms":106,"total_runtime_ms":200,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0241,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3339,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":282,"p95_latency_ms":344,"p99_latency_ms":369,"time_to_first_token_ms":102,"total_runtime_ms":313,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0284,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3677,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":259,"p95_latency_ms":316,"p99_latency_ms":339,"time_to_first_token_ms":107,"total_runtime_ms":287,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0109,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0831,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":244,"p95_latency_ms":298,"p99_latency_ms":319,"time_to_first_token_ms":96,"total_runtime_ms":271,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0274,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8017,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":186,"p95_latency_ms":227,"p99_latency_ms":244,"time_to_first_token_ms":76,"total_runtime_ms":206,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0145,"confidence_interval":0.9408,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2578,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":279,"p95_latency_ms":341,"p99_latency_ms":366,"time_to_first_token_ms":76,"total_runtime_ms":310,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0241,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1881,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":187,"p95_latency_ms":229,"p99_latency_ms":245,"time_to_first_token_ms":75,"total_runtime_ms":208,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0279,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0939,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":232,"p95_latency_ms":283,"p99_latency_ms":304,"time_to_first_token_ms":82,"total_runtime_ms":257,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0271,"confidence_interval":0.9481,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2582,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":660,"p95_latency_ms":807,"p99_latency_ms":866,"time_to_first_token_ms":167,"total_runtime_ms":734,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0263,"confidence_interval":0.9324,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.404,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":197,"total_runtime_ms":764,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0214,"confidence_interval":0.9401,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5863,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":618,"p95_latency_ms":756,"p99_latency_ms":811,"time_to_first_token_ms":166,"total_runtime_ms":687,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0109,"confidence_interval":0.9329,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0964,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":600,"p95_latency_ms":733,"p99_latency_ms":786,"time_to_first_token_ms":167,"total_runtime_ms":666,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0294,"confidence_interval":0.9372,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4734,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":674,"p95_latency_ms":824,"p99_latency_ms":884,"time_to_first_token_ms":194,"total_runtime_ms":749,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.023,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1049,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":703,"p95_latency_ms":859,"p99_latency_ms":921,"time_to_first_token_ms":163,"total_runtime_ms":781,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0122,"confidence_interval":0.932,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0098,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":607,"p95_latency_ms":742,"p99_latency_ms":796,"time_to_first_token_ms":173,"total_runtime_ms":674,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.148,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":625,"p95_latency_ms":764,"p99_latency_ms":820,"time_to_first_token_ms":195,"total_runtime_ms":695,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0165,"confidence_interval":0.9375,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2691,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":702,"p95_latency_ms":858,"p99_latency_ms":920,"time_to_first_token_ms":178,"total_runtime_ms":780,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0283,"confidence_interval":0.9305,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.7137,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":851,"time_to_first_token_ms":179,"total_runtime_ms":721,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0292,"confidence_interval":0.9361,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.86,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":345,"p95_latency_ms":422,"p99_latency_ms":452,"time_to_first_token_ms":127,"total_runtime_ms":383,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0121,"confidence_interval":0.9396,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8787,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":316,"p95_latency_ms":386,"p99_latency_ms":414,"time_to_first_token_ms":150,"total_runtime_ms":351,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0145,"confidence_interval":0.94,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1612,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":313,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":131,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0164,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4002,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":394,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":118,"total_runtime_ms":438,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0252,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1911,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":380,"p95_latency_ms":464,"p99_latency_ms":498,"time_to_first_token_ms":139,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0255,"confidence_interval":0.9381,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0478,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":513,"time_to_first_token_ms":136,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0213,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4763,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":339,"p95_latency_ms":414,"p99_latency_ms":444,"time_to_first_token_ms":137,"total_runtime_ms":377,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0269,"confidence_interval":0.9441,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9438,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":320,"p95_latency_ms":391,"p99_latency_ms":419,"time_to_first_token_ms":122,"total_runtime_ms":355,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0252,"confidence_interval":0.941,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4559,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":408,"time_to_first_token_ms":146,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0266,"confidence_interval":0.9492,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0065,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_minimal","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":407,"time_to_first_token_ms":147,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0147,"confidence_interval":0.9472,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1517,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":854,"p95_latency_ms":1044,"p99_latency_ms":1120,"time_to_first_token_ms":192,"total_runtime_ms":949,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6138,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":932,"p95_latency_ms":1139,"p99_latency_ms":1221,"time_to_first_token_ms":219,"total_runtime_ms":1035,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0183,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5223,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":899,"p95_latency_ms":1098,"p99_latency_ms":1178,"time_to_first_token_ms":213,"total_runtime_ms":999,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0188,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.169,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":874,"p95_latency_ms":1068,"p99_latency_ms":1146,"time_to_first_token_ms":191,"total_runtime_ms":971,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0299,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.82,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1014,"p99_latency_ms":1088,"time_to_first_token_ms":191,"total_runtime_ms":922,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0163,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4033,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":865,"p95_latency_ms":1057,"p99_latency_ms":1134,"time_to_first_token_ms":205,"total_runtime_ms":961,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0247,"confidence_interval":0.9434,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9527,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":922,"p95_latency_ms":1127,"p99_latency_ms":1209,"time_to_first_token_ms":197,"total_runtime_ms":1024,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0105,"confidence_interval":0.9445,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0845,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":883,"p95_latency_ms":1079,"p99_latency_ms":1158,"time_to_first_token_ms":203,"total_runtime_ms":981,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0287,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7913,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":902,"p95_latency_ms":1102,"p99_latency_ms":1182,"time_to_first_token_ms":209,"total_runtime_ms":1002,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0212,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8556,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":859,"p95_latency_ms":1050,"p99_latency_ms":1126,"time_to_first_token_ms":191,"total_runtime_ms":954,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.012,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5518,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":535,"p95_latency_ms":654,"p99_latency_ms":701,"time_to_first_token_ms":172,"total_runtime_ms":594,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4159,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":542,"p95_latency_ms":663,"p99_latency_ms":711,"time_to_first_token_ms":147,"total_runtime_ms":603,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0176,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9324,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":450,"p95_latency_ms":550,"p99_latency_ms":590,"time_to_first_token_ms":156,"total_runtime_ms":500,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0132,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.183,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":500,"p95_latency_ms":611,"p99_latency_ms":656,"time_to_first_token_ms":159,"total_runtime_ms":556,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0128,"confidence_interval":0.9438,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6676,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":473,"p95_latency_ms":578,"p99_latency_ms":621,"time_to_first_token_ms":171,"total_runtime_ms":526,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0206,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2874,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":514,"p95_latency_ms":628,"p99_latency_ms":674,"time_to_first_token_ms":171,"total_runtime_ms":571,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0168,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2711,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":552,"p99_latency_ms":592,"time_to_first_token_ms":159,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0157,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0944,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":174,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0153,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4382,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":461,"p95_latency_ms":563,"p99_latency_ms":604,"time_to_first_token_ms":151,"total_runtime_ms":512,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.021,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8977,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":467,"p95_latency_ms":570,"p99_latency_ms":612,"time_to_first_token_ms":173,"total_runtime_ms":518,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0162,"confidence_interval":0.9414,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9862,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1495,"p95_latency_ms":1827,"p99_latency_ms":1959,"time_to_first_token_ms":290,"total_runtime_ms":1661,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0283,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5182,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1450,"p95_latency_ms":1772,"p99_latency_ms":1901,"time_to_first_token_ms":281,"total_runtime_ms":1611,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0119,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9521,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1479,"p95_latency_ms":1808,"p99_latency_ms":1939,"time_to_first_token_ms":291,"total_runtime_ms":1644,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0209,"confidence_interval":0.943,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1532,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1503,"p95_latency_ms":1837,"p99_latency_ms":1970,"time_to_first_token_ms":288,"total_runtime_ms":1670,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0206,"confidence_interval":0.935,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5814,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1454,"p95_latency_ms":1777,"p99_latency_ms":1906,"time_to_first_token_ms":300,"total_runtime_ms":1616,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0189,"confidence_interval":0.9456,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3632,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1451,"p95_latency_ms":1773,"p99_latency_ms":1902,"time_to_first_token_ms":308,"total_runtime_ms":1612,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0215,"confidence_interval":0.9439,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5776,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1441,"p95_latency_ms":1761,"p99_latency_ms":1889,"time_to_first_token_ms":301,"total_runtime_ms":1601,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0125,"confidence_interval":0.9331,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6069,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1420,"p95_latency_ms":1736,"p99_latency_ms":1862,"time_to_first_token_ms":298,"total_runtime_ms":1578,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0126,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0555,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1440,"p95_latency_ms":1760,"p99_latency_ms":1888,"time_to_first_token_ms":302,"total_runtime_ms":1600,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0213,"confidence_interval":0.9419,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2723,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1492,"p95_latency_ms":1824,"p99_latency_ms":1956,"time_to_first_token_ms":294,"total_runtime_ms":1658,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0197,"confidence_interval":0.9322,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2822,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":858,"p95_latency_ms":1049,"p99_latency_ms":1125,"time_to_first_token_ms":234,"total_runtime_ms":954,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0178,"confidence_interval":0.9345,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5387,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1003,"time_to_first_token_ms":240,"total_runtime_ms":850,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.937,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2494,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":783,"p95_latency_ms":957,"p99_latency_ms":1027,"time_to_first_token_ms":257,"total_runtime_ms":870,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0288,"confidence_interval":0.9375,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6321,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":834,"p95_latency_ms":1020,"p99_latency_ms":1094,"time_to_first_token_ms":235,"total_runtime_ms":927,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0293,"confidence_interval":0.9331,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6825,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":858,"p95_latency_ms":1049,"p99_latency_ms":1125,"time_to_first_token_ms":235,"total_runtime_ms":954,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0257,"confidence_interval":0.933,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6513,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":830,"p95_latency_ms":1014,"p99_latency_ms":1088,"time_to_first_token_ms":267,"total_runtime_ms":922,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0117,"confidence_interval":0.9467,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3095,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":848,"p95_latency_ms":1036,"p99_latency_ms":1112,"time_to_first_token_ms":234,"total_runtime_ms":942,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.013,"confidence_interval":0.939,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9314,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":769,"p95_latency_ms":940,"p99_latency_ms":1008,"time_to_first_token_ms":249,"total_runtime_ms":854,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0165,"confidence_interval":0.9449,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2328,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":818,"p95_latency_ms":999,"p99_latency_ms":1072,"time_to_first_token_ms":238,"total_runtime_ms":909,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0222,"confidence_interval":0.9409,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8468,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":785,"p95_latency_ms":959,"p99_latency_ms":1029,"time_to_first_token_ms":257,"total_runtime_ms":872,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.024,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5439,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":643,"p95_latency_ms":785,"p99_latency_ms":842,"time_to_first_token_ms":162,"total_runtime_ms":714,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0175,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4388,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":683,"p95_latency_ms":834,"p99_latency_ms":895,"time_to_first_token_ms":178,"total_runtime_ms":758,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0114,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.276,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":701,"p95_latency_ms":856,"p99_latency_ms":919,"time_to_first_token_ms":177,"total_runtime_ms":779,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0188,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0499,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":673,"p95_latency_ms":823,"p99_latency_ms":882,"time_to_first_token_ms":157,"total_runtime_ms":748,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0264,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6745,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":624,"p95_latency_ms":762,"p99_latency_ms":818,"time_to_first_token_ms":168,"total_runtime_ms":693,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0249,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7669,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":627,"p95_latency_ms":766,"p99_latency_ms":822,"time_to_first_token_ms":175,"total_runtime_ms":696,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.022,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7038,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":697,"p95_latency_ms":852,"p99_latency_ms":913,"time_to_first_token_ms":165,"total_runtime_ms":774,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0295,"confidence_interval":0.9381,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8423,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":851,"time_to_first_token_ms":180,"total_runtime_ms":721,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0161,"confidence_interval":0.9336,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2339,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":949,"time_to_first_token_ms":173,"total_runtime_ms":804,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0153,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1774,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":637,"p95_latency_ms":779,"p99_latency_ms":835,"time_to_first_token_ms":159,"total_runtime_ms":708,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0184,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1477,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":339,"p95_latency_ms":414,"p99_latency_ms":444,"time_to_first_token_ms":118,"total_runtime_ms":376,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0287,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9535,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":384,"p95_latency_ms":469,"p99_latency_ms":504,"time_to_first_token_ms":136,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0169,"confidence_interval":0.9329,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6121,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":402,"p95_latency_ms":491,"p99_latency_ms":527,"time_to_first_token_ms":122,"total_runtime_ms":447,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0144,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.812,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":476,"p99_latency_ms":510,"time_to_first_token_ms":125,"total_runtime_ms":433,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0174,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1257,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":393,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":144,"total_runtime_ms":437,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.025,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6108,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":356,"p95_latency_ms":435,"p99_latency_ms":466,"time_to_first_token_ms":131,"total_runtime_ms":395,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0207,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7726,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":428,"p95_latency_ms":524,"p99_latency_ms":562,"time_to_first_token_ms":143,"total_runtime_ms":476,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.025,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0784,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":350,"p95_latency_ms":427,"p99_latency_ms":458,"time_to_first_token_ms":115,"total_runtime_ms":388,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.014,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6148,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":131,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0187,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5984,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":513,"time_to_first_token_ms":120,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0216,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6023,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1133,"p95_latency_ms":1384,"p99_latency_ms":1485,"time_to_first_token_ms":250,"total_runtime_ms":1259,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0275,"confidence_interval":0.9462,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9652,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1105,"p95_latency_ms":1351,"p99_latency_ms":1449,"time_to_first_token_ms":240,"total_runtime_ms":1228,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0257,"confidence_interval":0.9492,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2552,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1070,"p95_latency_ms":1308,"p99_latency_ms":1403,"time_to_first_token_ms":228,"total_runtime_ms":1189,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0248,"confidence_interval":0.9305,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5271,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1065,"p95_latency_ms":1302,"p99_latency_ms":1397,"time_to_first_token_ms":254,"total_runtime_ms":1184,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0213,"confidence_interval":0.9499,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9571,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1090,"p95_latency_ms":1332,"p99_latency_ms":1429,"time_to_first_token_ms":248,"total_runtime_ms":1211,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0216,"confidence_interval":0.9374,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8115,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1071,"p95_latency_ms":1309,"p99_latency_ms":1404,"time_to_first_token_ms":256,"total_runtime_ms":1190,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0189,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3746,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1067,"p95_latency_ms":1304,"p99_latency_ms":1398,"time_to_first_token_ms":231,"total_runtime_ms":1185,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0145,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.73,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1121,"p95_latency_ms":1370,"p99_latency_ms":1470,"time_to_first_token_ms":224,"total_runtime_ms":1246,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0129,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1274,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1090,"p95_latency_ms":1332,"p99_latency_ms":1429,"time_to_first_token_ms":245,"total_runtime_ms":1211,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0259,"confidence_interval":0.938,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8296,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1102,"p95_latency_ms":1346,"p99_latency_ms":1444,"time_to_first_token_ms":237,"total_runtime_ms":1224,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0277,"confidence_interval":0.9324,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6903,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":677,"p95_latency_ms":827,"p99_latency_ms":888,"time_to_first_token_ms":179,"total_runtime_ms":752,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0202,"confidence_interval":0.9406,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4596,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":610,"p95_latency_ms":746,"p99_latency_ms":800,"time_to_first_token_ms":202,"total_runtime_ms":678,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0217,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6902,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":629,"p95_latency_ms":769,"p99_latency_ms":825,"time_to_first_token_ms":199,"total_runtime_ms":699,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0222,"confidence_interval":0.9381,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2739,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":587,"p95_latency_ms":718,"p99_latency_ms":770,"time_to_first_token_ms":196,"total_runtime_ms":652,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0177,"confidence_interval":0.9462,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6113,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":646,"p95_latency_ms":790,"p99_latency_ms":847,"time_to_first_token_ms":202,"total_runtime_ms":718,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0125,"confidence_interval":0.9389,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5368,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":670,"p95_latency_ms":819,"p99_latency_ms":878,"time_to_first_token_ms":195,"total_runtime_ms":744,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0139,"confidence_interval":0.9412,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5276,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":676,"p95_latency_ms":826,"p99_latency_ms":886,"time_to_first_token_ms":195,"total_runtime_ms":751,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0238,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8539,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":625,"p95_latency_ms":764,"p99_latency_ms":819,"time_to_first_token_ms":179,"total_runtime_ms":694,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0155,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5157,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":652,"p95_latency_ms":797,"p99_latency_ms":855,"time_to_first_token_ms":188,"total_runtime_ms":724,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0237,"confidence_interval":0.9355,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4407,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":609,"p95_latency_ms":744,"p99_latency_ms":798,"time_to_first_token_ms":191,"total_runtime_ms":677,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0148,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0343,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":511,"p95_latency_ms":625,"p99_latency_ms":670,"time_to_first_token_ms":145,"total_runtime_ms":568,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.026,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5989,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":534,"p95_latency_ms":653,"p99_latency_ms":700,"time_to_first_token_ms":133,"total_runtime_ms":593,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0132,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.856,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":520,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":142,"total_runtime_ms":578,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0264,"confidence_interval":0.9452,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1169,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":545,"p95_latency_ms":666,"p99_latency_ms":714,"time_to_first_token_ms":141,"total_runtime_ms":605,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0211,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9172,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":565,"p95_latency_ms":691,"p99_latency_ms":741,"time_to_first_token_ms":154,"total_runtime_ms":628,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0229,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3164,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":464,"p95_latency_ms":568,"p99_latency_ms":609,"time_to_first_token_ms":164,"total_runtime_ms":516,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0181,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5301,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":548,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":136,"total_runtime_ms":609,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0296,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.438,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":561,"p95_latency_ms":686,"p99_latency_ms":736,"time_to_first_token_ms":164,"total_runtime_ms":623,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0261,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5911,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":145,"total_runtime_ms":576,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0234,"confidence_interval":0.9477,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5842,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":552,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":148,"total_runtime_ms":613,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0169,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8676,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":425,"time_to_first_token_ms":95,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0244,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7377,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":264,"p95_latency_ms":323,"p99_latency_ms":346,"time_to_first_token_ms":98,"total_runtime_ms":293,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0241,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2014,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":294,"p95_latency_ms":359,"p99_latency_ms":385,"time_to_first_token_ms":89,"total_runtime_ms":326,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0191,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9324,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":332,"p95_latency_ms":406,"p99_latency_ms":436,"time_to_first_token_ms":122,"total_runtime_ms":369,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0215,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3631,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":347,"p95_latency_ms":424,"p99_latency_ms":455,"time_to_first_token_ms":117,"total_runtime_ms":385,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0119,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0041,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":261,"p95_latency_ms":320,"p99_latency_ms":343,"time_to_first_token_ms":120,"total_runtime_ms":291,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5426,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":270,"p95_latency_ms":330,"p99_latency_ms":353,"time_to_first_token_ms":105,"total_runtime_ms":300,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0144,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.45,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":295,"p95_latency_ms":360,"p99_latency_ms":386,"time_to_first_token_ms":113,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0246,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9143,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":350,"p95_latency_ms":428,"p99_latency_ms":459,"time_to_first_token_ms":103,"total_runtime_ms":389,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0241,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3409,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":360,"p95_latency_ms":440,"p99_latency_ms":472,"time_to_first_token_ms":121,"total_runtime_ms":400,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0158,"confidence_interval":0.9307,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.07,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":795,"p95_latency_ms":971,"p99_latency_ms":1042,"time_to_first_token_ms":184,"total_runtime_ms":883,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0149,"confidence_interval":0.9492,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1523,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":801,"p95_latency_ms":979,"p99_latency_ms":1050,"time_to_first_token_ms":196,"total_runtime_ms":890,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0129,"confidence_interval":0.9435,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8185,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":859,"p95_latency_ms":1050,"p99_latency_ms":1127,"time_to_first_token_ms":186,"total_runtime_ms":955,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0144,"confidence_interval":0.9408,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.984,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":865,"p95_latency_ms":1058,"p99_latency_ms":1135,"time_to_first_token_ms":204,"total_runtime_ms":961,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0114,"confidence_interval":0.9443,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1994,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":832,"p95_latency_ms":1017,"p99_latency_ms":1091,"time_to_first_token_ms":183,"total_runtime_ms":925,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.016,"confidence_interval":0.9399,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5477,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":803,"p95_latency_ms":982,"p99_latency_ms":1053,"time_to_first_token_ms":185,"total_runtime_ms":893,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0125,"confidence_interval":0.9367,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0906,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":841,"p95_latency_ms":1027,"p99_latency_ms":1102,"time_to_first_token_ms":208,"total_runtime_ms":934,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0226,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.132,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":854,"p95_latency_ms":1043,"p99_latency_ms":1119,"time_to_first_token_ms":182,"total_runtime_ms":949,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9742,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":889,"p95_latency_ms":1087,"p99_latency_ms":1166,"time_to_first_token_ms":206,"total_runtime_ms":988,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0118,"confidence_interval":0.9469,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6839,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":893,"p95_latency_ms":1092,"p99_latency_ms":1171,"time_to_first_token_ms":197,"total_runtime_ms":992,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0116,"confidence_interval":0.9335,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5838,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":497,"p95_latency_ms":607,"p99_latency_ms":651,"time_to_first_token_ms":149,"total_runtime_ms":552,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0184,"confidence_interval":0.9392,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2588,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":591,"time_to_first_token_ms":169,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0176,"confidence_interval":0.9427,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9986,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":466,"p95_latency_ms":569,"p99_latency_ms":610,"time_to_first_token_ms":156,"total_runtime_ms":517,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0194,"confidence_interval":0.9338,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4402,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":456,"p95_latency_ms":557,"p99_latency_ms":598,"time_to_first_token_ms":150,"total_runtime_ms":506,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0153,"confidence_interval":0.9455,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2129,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":172,"total_runtime_ms":574,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0141,"confidence_interval":0.9415,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5472,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":480,"p95_latency_ms":587,"p99_latency_ms":629,"time_to_first_token_ms":150,"total_runtime_ms":533,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.03,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5304,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":504,"p95_latency_ms":617,"p99_latency_ms":661,"time_to_first_token_ms":159,"total_runtime_ms":560,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0225,"confidence_interval":0.9458,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4639,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":443,"p95_latency_ms":542,"p99_latency_ms":581,"time_to_first_token_ms":149,"total_runtime_ms":492,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8062,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":436,"p95_latency_ms":533,"p99_latency_ms":572,"time_to_first_token_ms":158,"total_runtime_ms":485,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0139,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9619,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":166,"total_runtime_ms":586,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0113,"confidence_interval":0.9471,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.124,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":133,"total_runtime_ms":510,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0147,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3036,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":142,"total_runtime_ms":510,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0277,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0268,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":414,"p95_latency_ms":506,"p99_latency_ms":543,"time_to_first_token_ms":142,"total_runtime_ms":460,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.023,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1301,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":490,"p95_latency_ms":598,"p99_latency_ms":642,"time_to_first_token_ms":142,"total_runtime_ms":544,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0196,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7286,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":435,"p95_latency_ms":532,"p99_latency_ms":571,"time_to_first_token_ms":130,"total_runtime_ms":483,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0159,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6517,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":424,"p95_latency_ms":518,"p99_latency_ms":556,"time_to_first_token_ms":128,"total_runtime_ms":471,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0227,"confidence_interval":0.9481,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4179,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":491,"p95_latency_ms":601,"p99_latency_ms":644,"time_to_first_token_ms":123,"total_runtime_ms":546,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0214,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7047,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":463,"p95_latency_ms":566,"p99_latency_ms":608,"time_to_first_token_ms":142,"total_runtime_ms":515,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3693,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":420,"p95_latency_ms":514,"p99_latency_ms":551,"time_to_first_token_ms":143,"total_runtime_ms":467,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0235,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8327,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":421,"p95_latency_ms":515,"p99_latency_ms":552,"time_to_first_token_ms":154,"total_runtime_ms":468,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0155,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.941,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":211,"p95_latency_ms":258,"p99_latency_ms":277,"time_to_first_token_ms":98,"total_runtime_ms":235,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0216,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1914,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":215,"p95_latency_ms":263,"p99_latency_ms":282,"time_to_first_token_ms":98,"total_runtime_ms":239,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0156,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1657,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":239,"p95_latency_ms":292,"p99_latency_ms":313,"time_to_first_token_ms":101,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0262,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4691,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":239,"p95_latency_ms":292,"p99_latency_ms":313,"time_to_first_token_ms":78,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0156,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5778,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":407,"time_to_first_token_ms":93,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0126,"confidence_interval":0.9455,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2523,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":281,"p95_latency_ms":343,"p99_latency_ms":368,"time_to_first_token_ms":99,"total_runtime_ms":312,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0119,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1679,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":317,"p95_latency_ms":387,"p99_latency_ms":415,"time_to_first_token_ms":87,"total_runtime_ms":352,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0111,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1873,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":244,"p95_latency_ms":299,"p99_latency_ms":320,"time_to_first_token_ms":95,"total_runtime_ms":271,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0118,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4907,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":304,"p95_latency_ms":371,"p99_latency_ms":398,"time_to_first_token_ms":96,"total_runtime_ms":338,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0271,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.149,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":211,"p95_latency_ms":258,"p99_latency_ms":277,"time_to_first_token_ms":109,"total_runtime_ms":235,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0173,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5518,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":720,"p95_latency_ms":881,"p99_latency_ms":945,"time_to_first_token_ms":176,"total_runtime_ms":800,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0157,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2053,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":703,"p95_latency_ms":860,"p99_latency_ms":922,"time_to_first_token_ms":188,"total_runtime_ms":781,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0233,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5478,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":763,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":168,"total_runtime_ms":848,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0236,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1282,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":724,"p95_latency_ms":884,"p99_latency_ms":949,"time_to_first_token_ms":175,"total_runtime_ms":804,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9338,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8403,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":705,"p95_latency_ms":862,"p99_latency_ms":924,"time_to_first_token_ms":178,"total_runtime_ms":783,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0241,"confidence_interval":0.9333,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2975,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":719,"p95_latency_ms":879,"p99_latency_ms":943,"time_to_first_token_ms":194,"total_runtime_ms":799,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0136,"confidence_interval":0.9367,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.288,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":735,"p95_latency_ms":899,"p99_latency_ms":964,"time_to_first_token_ms":179,"total_runtime_ms":817,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0297,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2946,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":167,"total_runtime_ms":741,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0164,"confidence_interval":0.9492,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1716,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":735,"p95_latency_ms":898,"p99_latency_ms":963,"time_to_first_token_ms":188,"total_runtime_ms":816,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7523,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":931,"time_to_first_token_ms":169,"total_runtime_ms":789,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0148,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0386,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":128,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0276,"confidence_interval":0.9359,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8677,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":376,"p95_latency_ms":459,"p99_latency_ms":492,"time_to_first_token_ms":147,"total_runtime_ms":417,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0277,"confidence_interval":0.9324,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.85,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":139,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0238,"confidence_interval":0.9466,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9017,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":119,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0132,"confidence_interval":0.9379,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4794,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":456,"p95_latency_ms":558,"p99_latency_ms":599,"time_to_first_token_ms":140,"total_runtime_ms":507,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0173,"confidence_interval":0.9475,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8917,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":137,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0143,"confidence_interval":0.9475,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.306,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":424,"p95_latency_ms":519,"p99_latency_ms":556,"time_to_first_token_ms":144,"total_runtime_ms":471,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0272,"confidence_interval":0.9352,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5429,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":381,"p95_latency_ms":466,"p99_latency_ms":500,"time_to_first_token_ms":118,"total_runtime_ms":424,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0155,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3002,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":453,"p95_latency_ms":554,"p99_latency_ms":594,"time_to_first_token_ms":132,"total_runtime_ms":503,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0285,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9156,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"rewrite_constrained","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":381,"p95_latency_ms":466,"p99_latency_ms":500,"time_to_first_token_ms":130,"total_runtime_ms":423,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0142,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4117,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1015,"p99_latency_ms":1089,"time_to_first_token_ms":200,"total_runtime_ms":923,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2041,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":816,"p95_latency_ms":998,"p99_latency_ms":1070,"time_to_first_token_ms":190,"total_runtime_ms":907,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0206,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0677,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":775,"p95_latency_ms":947,"p99_latency_ms":1016,"time_to_first_token_ms":199,"total_runtime_ms":861,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.021,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2216,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":764,"p95_latency_ms":934,"p99_latency_ms":1002,"time_to_first_token_ms":198,"total_runtime_ms":849,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0166,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3069,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":778,"p95_latency_ms":950,"p99_latency_ms":1019,"time_to_first_token_ms":212,"total_runtime_ms":864,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.03,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0329,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":210,"total_runtime_ms":889,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0289,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0653,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":811,"p95_latency_ms":991,"p99_latency_ms":1064,"time_to_first_token_ms":197,"total_runtime_ms":901,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0135,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9976,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":774,"p95_latency_ms":946,"p99_latency_ms":1015,"time_to_first_token_ms":192,"total_runtime_ms":860,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0187,"confidence_interval":0.9468,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5657,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":798,"p95_latency_ms":975,"p99_latency_ms":1046,"time_to_first_token_ms":190,"total_runtime_ms":887,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0281,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4724,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":789,"p95_latency_ms":965,"p99_latency_ms":1035,"time_to_first_token_ms":211,"total_runtime_ms":877,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0184,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2075,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":144,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0169,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5069,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":173,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0182,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4233,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":394,"p95_latency_ms":482,"p99_latency_ms":517,"time_to_first_token_ms":171,"total_runtime_ms":438,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.021,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4813,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":421,"p95_latency_ms":514,"p99_latency_ms":552,"time_to_first_token_ms":148,"total_runtime_ms":467,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0236,"confidence_interval":0.9317,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.272,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":479,"p95_latency_ms":585,"p99_latency_ms":628,"time_to_first_token_ms":169,"total_runtime_ms":532,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0195,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.28,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":406,"p95_latency_ms":497,"p99_latency_ms":533,"time_to_first_token_ms":157,"total_runtime_ms":452,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0101,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4927,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":448,"p95_latency_ms":547,"p99_latency_ms":587,"time_to_first_token_ms":162,"total_runtime_ms":498,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0245,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3865,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":393,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":162,"total_runtime_ms":437,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0172,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3368,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":581,"p99_latency_ms":623,"time_to_first_token_ms":167,"total_runtime_ms":528,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6648,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":416,"p95_latency_ms":509,"p99_latency_ms":546,"time_to_first_token_ms":158,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0159,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0109,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1299,"p95_latency_ms":1588,"p99_latency_ms":1703,"time_to_first_token_ms":295,"total_runtime_ms":1444,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0128,"confidence_interval":0.9335,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4457,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1286,"p95_latency_ms":1571,"p99_latency_ms":1686,"time_to_first_token_ms":301,"total_runtime_ms":1429,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.027,"confidence_interval":0.9408,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0404,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1345,"p95_latency_ms":1644,"p99_latency_ms":1764,"time_to_first_token_ms":296,"total_runtime_ms":1495,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.029,"confidence_interval":0.9445,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9954,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1343,"p95_latency_ms":1642,"p99_latency_ms":1761,"time_to_first_token_ms":290,"total_runtime_ms":1492,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0163,"confidence_interval":0.9442,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2736,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1300,"p95_latency_ms":1588,"p99_latency_ms":1704,"time_to_first_token_ms":295,"total_runtime_ms":1444,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0257,"confidence_interval":0.9434,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9125,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1358,"p95_latency_ms":1659,"p99_latency_ms":1780,"time_to_first_token_ms":292,"total_runtime_ms":1508,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0233,"confidence_interval":0.9393,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5587,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1342,"p95_latency_ms":1641,"p99_latency_ms":1760,"time_to_first_token_ms":304,"total_runtime_ms":1492,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0214,"confidence_interval":0.9375,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8051,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1330,"p95_latency_ms":1626,"p99_latency_ms":1744,"time_to_first_token_ms":312,"total_runtime_ms":1478,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0263,"confidence_interval":0.9406,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1218,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1326,"p95_latency_ms":1620,"p99_latency_ms":1738,"time_to_first_token_ms":282,"total_runtime_ms":1473,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0252,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1815,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1338,"p95_latency_ms":1635,"p99_latency_ms":1754,"time_to_first_token_ms":285,"total_runtime_ms":1487,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.023,"confidence_interval":0.9398,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9635,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":721,"p95_latency_ms":881,"p99_latency_ms":945,"time_to_first_token_ms":235,"total_runtime_ms":801,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0155,"confidence_interval":0.9381,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5919,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":700,"p95_latency_ms":856,"p99_latency_ms":918,"time_to_first_token_ms":258,"total_runtime_ms":778,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0288,"confidence_interval":0.9347,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5694,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":735,"p95_latency_ms":898,"p99_latency_ms":963,"time_to_first_token_ms":254,"total_runtime_ms":816,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.023,"confidence_interval":0.9493,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5766,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":650,"p95_latency_ms":794,"p99_latency_ms":852,"time_to_first_token_ms":237,"total_runtime_ms":722,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0191,"confidence_interval":0.9473,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2146,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":642,"p95_latency_ms":785,"p99_latency_ms":842,"time_to_first_token_ms":250,"total_runtime_ms":714,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0195,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2836,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":659,"p95_latency_ms":805,"p99_latency_ms":864,"time_to_first_token_ms":240,"total_runtime_ms":732,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0219,"confidence_interval":0.9337,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7554,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":682,"p95_latency_ms":834,"p99_latency_ms":894,"time_to_first_token_ms":240,"total_runtime_ms":758,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.026,"confidence_interval":0.9448,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.831,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":730,"p95_latency_ms":892,"p99_latency_ms":957,"time_to_first_token_ms":253,"total_runtime_ms":811,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0149,"confidence_interval":0.9478,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3524,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":686,"p95_latency_ms":839,"p99_latency_ms":900,"time_to_first_token_ms":261,"total_runtime_ms":762,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0286,"confidence_interval":0.9371,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2248,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":706,"p95_latency_ms":863,"p99_latency_ms":926,"time_to_first_token_ms":262,"total_runtime_ms":785,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.024,"confidence_interval":0.9447,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4144,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":601,"p95_latency_ms":735,"p99_latency_ms":788,"time_to_first_token_ms":172,"total_runtime_ms":668,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0188,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4274,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":756,"time_to_first_token_ms":163,"total_runtime_ms":641,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0142,"confidence_interval":0.944,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.542,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":568,"p95_latency_ms":694,"p99_latency_ms":745,"time_to_first_token_ms":163,"total_runtime_ms":631,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.025,"confidence_interval":0.9362,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7085,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":639,"p95_latency_ms":781,"p99_latency_ms":837,"time_to_first_token_ms":173,"total_runtime_ms":710,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0292,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2237,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":175,"total_runtime_ms":741,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0218,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8388,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":650,"p95_latency_ms":794,"p99_latency_ms":852,"time_to_first_token_ms":158,"total_runtime_ms":722,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0203,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5955,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":587,"p95_latency_ms":718,"p99_latency_ms":770,"time_to_first_token_ms":161,"total_runtime_ms":653,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0263,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6251,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":564,"p95_latency_ms":689,"p99_latency_ms":739,"time_to_first_token_ms":188,"total_runtime_ms":626,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0136,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1873,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":624,"p95_latency_ms":763,"p99_latency_ms":818,"time_to_first_token_ms":169,"total_runtime_ms":694,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0149,"confidence_interval":0.9481,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0833,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":645,"p95_latency_ms":788,"p99_latency_ms":846,"time_to_first_token_ms":181,"total_runtime_ms":717,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0185,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8327,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":325,"p95_latency_ms":398,"p99_latency_ms":426,"time_to_first_token_ms":141,"total_runtime_ms":361,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0185,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6295,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":343,"p95_latency_ms":419,"p99_latency_ms":450,"time_to_first_token_ms":131,"total_runtime_ms":381,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.019,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1968,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":479,"time_to_first_token_ms":114,"total_runtime_ms":406,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0107,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1281,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":329,"p95_latency_ms":402,"p99_latency_ms":431,"time_to_first_token_ms":122,"total_runtime_ms":365,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0153,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4114,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":336,"p95_latency_ms":410,"p99_latency_ms":440,"time_to_first_token_ms":132,"total_runtime_ms":373,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0181,"confidence_interval":0.9399,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1279,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":349,"p95_latency_ms":426,"p99_latency_ms":457,"time_to_first_token_ms":131,"total_runtime_ms":387,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0243,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.055,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":287,"p95_latency_ms":350,"p99_latency_ms":376,"time_to_first_token_ms":126,"total_runtime_ms":318,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0271,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7284,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":295,"p95_latency_ms":360,"p99_latency_ms":386,"time_to_first_token_ms":139,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.021,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6775,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":283,"p95_latency_ms":346,"p99_latency_ms":371,"time_to_first_token_ms":113,"total_runtime_ms":315,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0168,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5518,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":293,"p95_latency_ms":358,"p99_latency_ms":384,"time_to_first_token_ms":131,"total_runtime_ms":325,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0227,"confidence_interval":0.948,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8421,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":985,"p95_latency_ms":1204,"p99_latency_ms":1291,"time_to_first_token_ms":250,"total_runtime_ms":1094,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.013,"confidence_interval":0.9395,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3034,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1021,"p95_latency_ms":1248,"p99_latency_ms":1338,"time_to_first_token_ms":239,"total_runtime_ms":1134,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0223,"confidence_interval":0.9489,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3061,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":972,"p95_latency_ms":1188,"p99_latency_ms":1274,"time_to_first_token_ms":241,"total_runtime_ms":1080,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0131,"confidence_interval":0.9375,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1362,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":983,"p95_latency_ms":1201,"p99_latency_ms":1288,"time_to_first_token_ms":238,"total_runtime_ms":1092,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0191,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0781,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1003,"p95_latency_ms":1226,"p99_latency_ms":1315,"time_to_first_token_ms":245,"total_runtime_ms":1114,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0176,"confidence_interval":0.9499,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8757,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":976,"p95_latency_ms":1193,"p99_latency_ms":1279,"time_to_first_token_ms":250,"total_runtime_ms":1084,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0258,"confidence_interval":0.9436,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1365,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":956,"p95_latency_ms":1168,"p99_latency_ms":1253,"time_to_first_token_ms":231,"total_runtime_ms":1062,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0148,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4468,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1008,"p95_latency_ms":1232,"p99_latency_ms":1322,"time_to_first_token_ms":246,"total_runtime_ms":1120,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0296,"confidence_interval":0.9371,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7534,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":988,"p95_latency_ms":1207,"p99_latency_ms":1295,"time_to_first_token_ms":252,"total_runtime_ms":1097,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0106,"confidence_interval":0.9333,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5683,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1033,"p95_latency_ms":1262,"p99_latency_ms":1354,"time_to_first_token_ms":245,"total_runtime_ms":1148,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0164,"confidence_interval":0.9318,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1676,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":479,"p95_latency_ms":585,"p99_latency_ms":628,"time_to_first_token_ms":206,"total_runtime_ms":532,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0198,"confidence_interval":0.9461,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0633,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":503,"p95_latency_ms":615,"p99_latency_ms":660,"time_to_first_token_ms":193,"total_runtime_ms":559,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0206,"confidence_interval":0.9331,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1138,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":578,"p95_latency_ms":706,"p99_latency_ms":757,"time_to_first_token_ms":211,"total_runtime_ms":642,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0107,"confidence_interval":0.9383,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.375,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":514,"p95_latency_ms":628,"p99_latency_ms":674,"time_to_first_token_ms":208,"total_runtime_ms":571,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0154,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4468,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":547,"p95_latency_ms":668,"p99_latency_ms":717,"time_to_first_token_ms":203,"total_runtime_ms":607,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0271,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0054,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":503,"p95_latency_ms":614,"p99_latency_ms":659,"time_to_first_token_ms":205,"total_runtime_ms":559,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0123,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.27,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":532,"p95_latency_ms":650,"p99_latency_ms":697,"time_to_first_token_ms":191,"total_runtime_ms":591,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0139,"confidence_interval":0.9344,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2555,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":519,"p95_latency_ms":635,"p99_latency_ms":681,"time_to_first_token_ms":180,"total_runtime_ms":577,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0268,"confidence_interval":0.9453,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6193,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":522,"p95_latency_ms":638,"p99_latency_ms":684,"time_to_first_token_ms":190,"total_runtime_ms":580,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.02,"confidence_interval":0.9477,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6929,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":197,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0138,"confidence_interval":0.9385,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3127,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":479,"p95_latency_ms":585,"p99_latency_ms":628,"time_to_first_token_ms":155,"total_runtime_ms":532,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.022,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0602,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":147,"total_runtime_ms":586,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0213,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.442,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":500,"p95_latency_ms":611,"p99_latency_ms":655,"time_to_first_token_ms":159,"total_runtime_ms":555,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0161,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.14,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":522,"p95_latency_ms":638,"p99_latency_ms":684,"time_to_first_token_ms":147,"total_runtime_ms":580,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0133,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4957,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":495,"p95_latency_ms":605,"p99_latency_ms":649,"time_to_first_token_ms":150,"total_runtime_ms":550,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.014,"confidence_interval":0.938,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6037,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":680,"time_to_first_token_ms":157,"total_runtime_ms":576,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0203,"confidence_interval":0.9329,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1287,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":455,"p95_latency_ms":556,"p99_latency_ms":596,"time_to_first_token_ms":153,"total_runtime_ms":505,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0243,"confidence_interval":0.9327,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3347,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":529,"p95_latency_ms":647,"p99_latency_ms":694,"time_to_first_token_ms":156,"total_runtime_ms":588,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0167,"confidence_interval":0.9348,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7011,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":134,"total_runtime_ms":575,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.019,"confidence_interval":0.9451,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0742,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":528,"p95_latency_ms":645,"p99_latency_ms":692,"time_to_first_token_ms":145,"total_runtime_ms":586,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0123,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0475,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":376,"p99_latency_ms":404,"time_to_first_token_ms":100,"total_runtime_ms":342,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3725,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":314,"p95_latency_ms":384,"p99_latency_ms":412,"time_to_first_token_ms":94,"total_runtime_ms":349,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0175,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.728,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":406,"time_to_first_token_ms":98,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0149,"confidence_interval":0.9329,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8747,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":238,"p95_latency_ms":291,"p99_latency_ms":312,"time_to_first_token_ms":121,"total_runtime_ms":265,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0296,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0677,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":357,"p99_latency_ms":383,"time_to_first_token_ms":107,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0237,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8714,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":215,"p95_latency_ms":263,"p99_latency_ms":282,"time_to_first_token_ms":119,"total_runtime_ms":239,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0116,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1104,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":281,"p95_latency_ms":344,"p99_latency_ms":369,"time_to_first_token_ms":109,"total_runtime_ms":312,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0225,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1297,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":277,"p95_latency_ms":339,"p99_latency_ms":364,"time_to_first_token_ms":94,"total_runtime_ms":308,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0137,"confidence_interval":0.9307,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6733,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":251,"p95_latency_ms":307,"p99_latency_ms":329,"time_to_first_token_ms":98,"total_runtime_ms":279,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0289,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3719,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":241,"p95_latency_ms":295,"p99_latency_ms":316,"time_to_first_token_ms":117,"total_runtime_ms":268,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0148,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8718,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":744,"p95_latency_ms":910,"p99_latency_ms":976,"time_to_first_token_ms":212,"total_runtime_ms":827,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.013,"confidence_interval":0.9363,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.7081,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":768,"p95_latency_ms":939,"p99_latency_ms":1007,"time_to_first_token_ms":203,"total_runtime_ms":853,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0286,"confidence_interval":0.9418,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9588,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":812,"p95_latency_ms":993,"p99_latency_ms":1065,"time_to_first_token_ms":192,"total_runtime_ms":903,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0182,"confidence_interval":0.9329,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3268,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":791,"p95_latency_ms":967,"p99_latency_ms":1037,"time_to_first_token_ms":206,"total_runtime_ms":879,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0132,"confidence_interval":0.9368,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8038,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":765,"p95_latency_ms":935,"p99_latency_ms":1003,"time_to_first_token_ms":217,"total_runtime_ms":850,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0254,"confidence_interval":0.9405,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4778,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":751,"p95_latency_ms":918,"p99_latency_ms":985,"time_to_first_token_ms":196,"total_runtime_ms":834,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0163,"confidence_interval":0.9474,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5416,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":770,"p95_latency_ms":941,"p99_latency_ms":1010,"time_to_first_token_ms":191,"total_runtime_ms":856,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0158,"confidence_interval":0.9412,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7866,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":786,"p95_latency_ms":961,"p99_latency_ms":1031,"time_to_first_token_ms":209,"total_runtime_ms":873,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0139,"confidence_interval":0.9409,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4493,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":717,"p95_latency_ms":877,"p99_latency_ms":940,"time_to_first_token_ms":205,"total_runtime_ms":797,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0236,"confidence_interval":0.9356,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9204,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":747,"p95_latency_ms":913,"p99_latency_ms":979,"time_to_first_token_ms":210,"total_runtime_ms":830,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.014,"confidence_interval":0.9494,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1288,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":416,"p95_latency_ms":509,"p99_latency_ms":546,"time_to_first_token_ms":157,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0152,"confidence_interval":0.9422,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7898,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":454,"p95_latency_ms":554,"p99_latency_ms":595,"time_to_first_token_ms":161,"total_runtime_ms":504,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0175,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.57,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":507,"time_to_first_token_ms":154,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0144,"confidence_interval":0.9386,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9138,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":448,"p95_latency_ms":547,"p99_latency_ms":587,"time_to_first_token_ms":154,"total_runtime_ms":497,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0218,"confidence_interval":0.9327,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.241,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":496,"time_to_first_token_ms":142,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.019,"confidence_interval":0.9476,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3799,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":433,"p95_latency_ms":530,"p99_latency_ms":568,"time_to_first_token_ms":150,"total_runtime_ms":482,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0151,"confidence_interval":0.94,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.7027,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":413,"p95_latency_ms":505,"p99_latency_ms":542,"time_to_first_token_ms":148,"total_runtime_ms":459,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.9315,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.479,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":367,"p95_latency_ms":449,"p99_latency_ms":482,"time_to_first_token_ms":161,"total_runtime_ms":408,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0252,"confidence_interval":0.9316,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9516,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":375,"p95_latency_ms":458,"p99_latency_ms":491,"time_to_first_token_ms":146,"total_runtime_ms":416,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0213,"confidence_interval":0.9429,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4885,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":377,"p95_latency_ms":460,"p99_latency_ms":494,"time_to_first_token_ms":169,"total_runtime_ms":418,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.01,"confidence_interval":0.9418,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2182,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":382,"p95_latency_ms":467,"p99_latency_ms":501,"time_to_first_token_ms":138,"total_runtime_ms":424,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0212,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9032,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":434,"p95_latency_ms":531,"p99_latency_ms":570,"time_to_first_token_ms":141,"total_runtime_ms":483,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0137,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2848,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":457,"p95_latency_ms":559,"p99_latency_ms":600,"time_to_first_token_ms":132,"total_runtime_ms":508,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0214,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.255,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":470,"time_to_first_token_ms":120,"total_runtime_ms":399,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9553,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":425,"p95_latency_ms":519,"p99_latency_ms":557,"time_to_first_token_ms":143,"total_runtime_ms":472,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0254,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.957,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":426,"p95_latency_ms":521,"p99_latency_ms":559,"time_to_first_token_ms":153,"total_runtime_ms":473,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0278,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9016,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":577,"time_to_first_token_ms":130,"total_runtime_ms":489,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0247,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5806,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":399,"p95_latency_ms":488,"p99_latency_ms":523,"time_to_first_token_ms":122,"total_runtime_ms":443,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0184,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2662,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":382,"p95_latency_ms":467,"p99_latency_ms":501,"time_to_first_token_ms":152,"total_runtime_ms":424,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0294,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4137,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":482,"time_to_first_token_ms":152,"total_runtime_ms":409,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0117,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9823,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":252,"p95_latency_ms":307,"p99_latency_ms":330,"time_to_first_token_ms":80,"total_runtime_ms":279,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0273,"confidence_interval":0.942,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6272,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":249,"p95_latency_ms":305,"p99_latency_ms":327,"time_to_first_token_ms":109,"total_runtime_ms":277,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0267,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8709,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":271,"p95_latency_ms":331,"p99_latency_ms":355,"time_to_first_token_ms":85,"total_runtime_ms":301,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0107,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3136,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":282,"p95_latency_ms":345,"p99_latency_ms":370,"time_to_first_token_ms":88,"total_runtime_ms":313,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0177,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1736,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":281,"p95_latency_ms":343,"p99_latency_ms":368,"time_to_first_token_ms":100,"total_runtime_ms":312,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0103,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7176,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":189,"p95_latency_ms":231,"p99_latency_ms":248,"time_to_first_token_ms":81,"total_runtime_ms":210,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0283,"confidence_interval":0.944,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.989,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":258,"p95_latency_ms":315,"p99_latency_ms":338,"time_to_first_token_ms":99,"total_runtime_ms":287,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0117,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7569,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":225,"p95_latency_ms":275,"p99_latency_ms":295,"time_to_first_token_ms":104,"total_runtime_ms":250,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.027,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0718,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":216,"p95_latency_ms":264,"p99_latency_ms":283,"time_to_first_token_ms":89,"total_runtime_ms":240,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0228,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2043,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":251,"p95_latency_ms":306,"p99_latency_ms":329,"time_to_first_token_ms":103,"total_runtime_ms":279,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0265,"confidence_interval":0.9327,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6427,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":678,"p95_latency_ms":829,"p99_latency_ms":889,"time_to_first_token_ms":171,"total_runtime_ms":753,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0231,"confidence_interval":0.9457,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5231,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":690,"p95_latency_ms":843,"p99_latency_ms":904,"time_to_first_token_ms":185,"total_runtime_ms":766,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0219,"confidence_interval":0.9378,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1308,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":696,"p95_latency_ms":851,"p99_latency_ms":913,"time_to_first_token_ms":183,"total_runtime_ms":773,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0178,"confidence_interval":0.9487,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2462,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":614,"p95_latency_ms":750,"p99_latency_ms":804,"time_to_first_token_ms":191,"total_runtime_ms":682,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.014,"confidence_interval":0.9327,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2204,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":844,"time_to_first_token_ms":170,"total_runtime_ms":715,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0244,"confidence_interval":0.9425,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3587,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":651,"p95_latency_ms":796,"p99_latency_ms":854,"time_to_first_token_ms":176,"total_runtime_ms":724,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0282,"confidence_interval":0.9428,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3544,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":611,"p95_latency_ms":747,"p99_latency_ms":801,"time_to_first_token_ms":192,"total_runtime_ms":679,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0153,"confidence_interval":0.9461,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5384,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":699,"p95_latency_ms":854,"p99_latency_ms":916,"time_to_first_token_ms":174,"total_runtime_ms":777,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.029,"confidence_interval":0.9317,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9625,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":191,"total_runtime_ms":663,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0295,"confidence_interval":0.9369,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4223,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":679,"p95_latency_ms":830,"p99_latency_ms":890,"time_to_first_token_ms":191,"total_runtime_ms":754,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0226,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0116,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":344,"p95_latency_ms":420,"p99_latency_ms":451,"time_to_first_token_ms":128,"total_runtime_ms":382,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0227,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0869,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":122,"total_runtime_ms":436,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0212,"confidence_interval":0.9386,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4251,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":299,"p95_latency_ms":365,"p99_latency_ms":392,"time_to_first_token_ms":149,"total_runtime_ms":332,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.028,"confidence_interval":0.9394,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5583,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":513,"time_to_first_token_ms":131,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0221,"confidence_interval":0.9357,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2036,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":385,"p95_latency_ms":470,"p99_latency_ms":505,"time_to_first_token_ms":122,"total_runtime_ms":428,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.016,"confidence_interval":0.9467,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8759,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":395,"p95_latency_ms":483,"p99_latency_ms":518,"time_to_first_token_ms":125,"total_runtime_ms":439,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.01,"confidence_interval":0.9382,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8435,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":404,"p95_latency_ms":494,"p99_latency_ms":530,"time_to_first_token_ms":122,"total_runtime_ms":449,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0258,"confidence_interval":0.9393,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7618,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":331,"p95_latency_ms":405,"p99_latency_ms":434,"time_to_first_token_ms":134,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.025,"confidence_interval":0.9397,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6894,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":397,"p99_latency_ms":425,"time_to_first_token_ms":141,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.02,"confidence_interval":0.9335,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2221,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_binary|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_binary","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":506,"time_to_first_token_ms":146,"total_runtime_ms":429,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.407,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":894,"p95_latency_ms":1092,"p99_latency_ms":1172,"time_to_first_token_ms":201,"total_runtime_ms":993,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9115,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":911,"p95_latency_ms":1114,"p99_latency_ms":1195,"time_to_first_token_ms":211,"total_runtime_ms":1012,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0215,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0809,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":871,"p95_latency_ms":1065,"p99_latency_ms":1143,"time_to_first_token_ms":193,"total_runtime_ms":968,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0127,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9941,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":882,"p95_latency_ms":1078,"p99_latency_ms":1156,"time_to_first_token_ms":204,"total_runtime_ms":980,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0154,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3396,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":889,"p95_latency_ms":1087,"p99_latency_ms":1166,"time_to_first_token_ms":222,"total_runtime_ms":988,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0265,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9437,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":829,"p95_latency_ms":1014,"p99_latency_ms":1087,"time_to_first_token_ms":202,"total_runtime_ms":922,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0258,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1538,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":918,"p95_latency_ms":1122,"p99_latency_ms":1204,"time_to_first_token_ms":207,"total_runtime_ms":1020,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0191,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3812,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":873,"p95_latency_ms":1067,"p99_latency_ms":1144,"time_to_first_token_ms":193,"total_runtime_ms":970,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0268,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8663,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1030,"p99_latency_ms":1105,"time_to_first_token_ms":206,"total_runtime_ms":937,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0198,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7312,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":880,"p95_latency_ms":1075,"p99_latency_ms":1153,"time_to_first_token_ms":219,"total_runtime_ms":977,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0225,"confidence_interval":0.9416,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5039,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":525,"p95_latency_ms":641,"p99_latency_ms":688,"time_to_first_token_ms":147,"total_runtime_ms":583,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0133,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3849,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":486,"p95_latency_ms":594,"p99_latency_ms":638,"time_to_first_token_ms":165,"total_runtime_ms":540,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.015,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6904,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":493,"p95_latency_ms":602,"p99_latency_ms":646,"time_to_first_token_ms":164,"total_runtime_ms":548,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0297,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9214,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":592,"time_to_first_token_ms":164,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0196,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5524,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":472,"p95_latency_ms":576,"p99_latency_ms":618,"time_to_first_token_ms":160,"total_runtime_ms":524,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0249,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7072,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":471,"p95_latency_ms":575,"p99_latency_ms":617,"time_to_first_token_ms":152,"total_runtime_ms":523,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0255,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4965,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":552,"p95_latency_ms":675,"p99_latency_ms":724,"time_to_first_token_ms":157,"total_runtime_ms":614,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0138,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7362,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":493,"p95_latency_ms":602,"p99_latency_ms":646,"time_to_first_token_ms":152,"total_runtime_ms":547,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0115,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9877,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":476,"p95_latency_ms":581,"p99_latency_ms":624,"time_to_first_token_ms":161,"total_runtime_ms":528,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0165,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6559,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":522,"p95_latency_ms":638,"p99_latency_ms":685,"time_to_first_token_ms":169,"total_runtime_ms":580,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0192,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2625,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1493,"p95_latency_ms":1825,"p99_latency_ms":1958,"time_to_first_token_ms":293,"total_runtime_ms":1659,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.011,"confidence_interval":0.938,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4623,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1407,"p95_latency_ms":1720,"p99_latency_ms":1845,"time_to_first_token_ms":288,"total_runtime_ms":1564,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4578,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1505,"p95_latency_ms":1839,"p99_latency_ms":1973,"time_to_first_token_ms":310,"total_runtime_ms":1672,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0133,"confidence_interval":0.9476,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.36,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1439,"p95_latency_ms":1759,"p99_latency_ms":1887,"time_to_first_token_ms":284,"total_runtime_ms":1599,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0207,"confidence_interval":0.9437,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6434,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1426,"p95_latency_ms":1743,"p99_latency_ms":1870,"time_to_first_token_ms":298,"total_runtime_ms":1585,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.018,"confidence_interval":0.9474,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1708,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1406,"p95_latency_ms":1719,"p99_latency_ms":1844,"time_to_first_token_ms":293,"total_runtime_ms":1563,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0206,"confidence_interval":0.9416,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0158,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1502,"p95_latency_ms":1836,"p99_latency_ms":1970,"time_to_first_token_ms":311,"total_runtime_ms":1669,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0271,"confidence_interval":0.9342,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4589,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1400,"p95_latency_ms":1712,"p99_latency_ms":1836,"time_to_first_token_ms":311,"total_runtime_ms":1556,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0114,"confidence_interval":0.9445,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0388,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1475,"p95_latency_ms":1802,"p99_latency_ms":1933,"time_to_first_token_ms":312,"total_runtime_ms":1638,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0162,"confidence_interval":0.9328,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1621,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1495,"p95_latency_ms":1827,"p99_latency_ms":1960,"time_to_first_token_ms":281,"total_runtime_ms":1661,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":71,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0243,"confidence_interval":0.9462,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9503,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":775,"p95_latency_ms":947,"p99_latency_ms":1016,"time_to_first_token_ms":251,"total_runtime_ms":861,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0126,"confidence_interval":0.9401,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8551,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":849,"p95_latency_ms":1038,"p99_latency_ms":1113,"time_to_first_token_ms":263,"total_runtime_ms":943,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.026,"confidence_interval":0.9383,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0352,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":868,"p95_latency_ms":1060,"p99_latency_ms":1138,"time_to_first_token_ms":256,"total_runtime_ms":964,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0175,"confidence_interval":0.9418,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5489,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":853,"p95_latency_ms":1042,"p99_latency_ms":1118,"time_to_first_token_ms":241,"total_runtime_ms":947,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0135,"confidence_interval":0.9376,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2751,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1053,"time_to_first_token_ms":263,"total_runtime_ms":892,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0244,"confidence_interval":0.937,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4125,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":867,"p95_latency_ms":1060,"p99_latency_ms":1137,"time_to_first_token_ms":258,"total_runtime_ms":963,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0102,"confidence_interval":0.947,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9357,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":841,"p95_latency_ms":1028,"p99_latency_ms":1103,"time_to_first_token_ms":258,"total_runtime_ms":934,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0204,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.99,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":856,"p95_latency_ms":1047,"p99_latency_ms":1123,"time_to_first_token_ms":243,"total_runtime_ms":952,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0173,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.689,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":872,"p95_latency_ms":1065,"p99_latency_ms":1143,"time_to_first_token_ms":253,"total_runtime_ms":968,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0118,"confidence_interval":0.9341,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3124,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":789,"p95_latency_ms":965,"p99_latency_ms":1035,"time_to_first_token_ms":266,"total_runtime_ms":877,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0248,"confidence_interval":0.931,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.655,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":685,"p95_latency_ms":837,"p99_latency_ms":898,"time_to_first_token_ms":162,"total_runtime_ms":761,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0204,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2349,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":653,"p95_latency_ms":798,"p99_latency_ms":856,"time_to_first_token_ms":158,"total_runtime_ms":725,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.024,"confidence_interval":0.9317,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3939,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":843,"time_to_first_token_ms":175,"total_runtime_ms":715,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.013,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8897,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":158,"total_runtime_ms":765,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0204,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8094,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":692,"p95_latency_ms":845,"p99_latency_ms":907,"time_to_first_token_ms":190,"total_runtime_ms":769,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0245,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2441,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":679,"p95_latency_ms":830,"p99_latency_ms":890,"time_to_first_token_ms":175,"total_runtime_ms":755,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0267,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5258,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":696,"p95_latency_ms":850,"p99_latency_ms":912,"time_to_first_token_ms":168,"total_runtime_ms":773,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0203,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9148,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":716,"p95_latency_ms":876,"p99_latency_ms":939,"time_to_first_token_ms":170,"total_runtime_ms":796,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0201,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3578,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":725,"p95_latency_ms":886,"p99_latency_ms":951,"time_to_first_token_ms":185,"total_runtime_ms":806,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0101,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0064,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":701,"p95_latency_ms":857,"p99_latency_ms":919,"time_to_first_token_ms":187,"total_runtime_ms":779,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0155,"confidence_interval":0.9479,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3976,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":354,"p95_latency_ms":432,"p99_latency_ms":464,"time_to_first_token_ms":128,"total_runtime_ms":393,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0145,"confidence_interval":0.9347,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0938,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":513,"time_to_first_token_ms":113,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0153,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2116,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":434,"p95_latency_ms":531,"p99_latency_ms":569,"time_to_first_token_ms":116,"total_runtime_ms":482,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0153,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5224,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":471,"time_to_first_token_ms":115,"total_runtime_ms":399,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0124,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7233,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":411,"p95_latency_ms":503,"p99_latency_ms":539,"time_to_first_token_ms":129,"total_runtime_ms":457,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0158,"confidence_interval":0.942,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6793,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":437,"p99_latency_ms":469,"time_to_first_token_ms":120,"total_runtime_ms":398,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0195,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7414,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":471,"time_to_first_token_ms":144,"total_runtime_ms":399,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0156,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4219,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":357,"p95_latency_ms":437,"p99_latency_ms":468,"time_to_first_token_ms":122,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0143,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3185,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":398,"p95_latency_ms":486,"p99_latency_ms":522,"time_to_first_token_ms":121,"total_runtime_ms":442,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0195,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9641,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":401,"p95_latency_ms":490,"p99_latency_ms":525,"time_to_first_token_ms":131,"total_runtime_ms":445,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0295,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.203,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1068,"p95_latency_ms":1305,"p99_latency_ms":1400,"time_to_first_token_ms":246,"total_runtime_ms":1186,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9482,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6346,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1074,"p95_latency_ms":1313,"p99_latency_ms":1409,"time_to_first_token_ms":232,"total_runtime_ms":1194,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0293,"confidence_interval":0.9338,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0493,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1147,"p95_latency_ms":1402,"p99_latency_ms":1504,"time_to_first_token_ms":234,"total_runtime_ms":1275,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0162,"confidence_interval":0.935,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4641,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1081,"p95_latency_ms":1322,"p99_latency_ms":1418,"time_to_first_token_ms":254,"total_runtime_ms":1201,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0212,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9903,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1068,"p95_latency_ms":1305,"p99_latency_ms":1400,"time_to_first_token_ms":238,"total_runtime_ms":1186,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0266,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0216,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1141,"p95_latency_ms":1394,"p99_latency_ms":1495,"time_to_first_token_ms":241,"total_runtime_ms":1267,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.014,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9791,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1079,"p95_latency_ms":1318,"p99_latency_ms":1414,"time_to_first_token_ms":233,"total_runtime_ms":1198,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0124,"confidence_interval":0.9352,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9219,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1111,"p95_latency_ms":1358,"p99_latency_ms":1457,"time_to_first_token_ms":233,"total_runtime_ms":1235,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0183,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4773,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1065,"p95_latency_ms":1302,"p99_latency_ms":1397,"time_to_first_token_ms":255,"total_runtime_ms":1184,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0139,"confidence_interval":0.9315,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.202,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1125,"p95_latency_ms":1375,"p99_latency_ms":1475,"time_to_first_token_ms":257,"total_runtime_ms":1250,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":56,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0193,"confidence_interval":0.9361,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.722,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":593,"p95_latency_ms":724,"p99_latency_ms":777,"time_to_first_token_ms":209,"total_runtime_ms":659,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0209,"confidence_interval":0.9431,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4912,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":593,"p95_latency_ms":725,"p99_latency_ms":778,"time_to_first_token_ms":197,"total_runtime_ms":659,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0288,"confidence_interval":0.9309,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3275,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":673,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":204,"total_runtime_ms":748,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0265,"confidence_interval":0.9473,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3264,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":656,"p95_latency_ms":801,"p99_latency_ms":860,"time_to_first_token_ms":212,"total_runtime_ms":728,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0152,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.958,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":618,"p95_latency_ms":755,"p99_latency_ms":810,"time_to_first_token_ms":182,"total_runtime_ms":686,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0136,"confidence_interval":0.9395,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5069,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":660,"p95_latency_ms":807,"p99_latency_ms":866,"time_to_first_token_ms":196,"total_runtime_ms":734,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6993,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":623,"p95_latency_ms":762,"p99_latency_ms":817,"time_to_first_token_ms":204,"total_runtime_ms":693,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0222,"confidence_interval":0.9348,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0864,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":803,"time_to_first_token_ms":195,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0294,"confidence_interval":0.9343,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4361,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":660,"p95_latency_ms":806,"p99_latency_ms":865,"time_to_first_token_ms":182,"total_runtime_ms":733,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0143,"confidence_interval":0.9364,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5079,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":627,"p95_latency_ms":767,"p99_latency_ms":823,"time_to_first_token_ms":191,"total_runtime_ms":697,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9402,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3436,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":541,"p95_latency_ms":661,"p99_latency_ms":709,"time_to_first_token_ms":143,"total_runtime_ms":601,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5673,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":499,"p95_latency_ms":610,"p99_latency_ms":655,"time_to_first_token_ms":147,"total_runtime_ms":555,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0207,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4005,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":552,"p95_latency_ms":675,"p99_latency_ms":724,"time_to_first_token_ms":141,"total_runtime_ms":613,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0156,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.306,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":483,"p95_latency_ms":590,"p99_latency_ms":633,"time_to_first_token_ms":160,"total_runtime_ms":536,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0227,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1138,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":556,"p95_latency_ms":680,"p99_latency_ms":729,"time_to_first_token_ms":161,"total_runtime_ms":618,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0174,"confidence_interval":0.93,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8259,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":509,"p95_latency_ms":622,"p99_latency_ms":667,"time_to_first_token_ms":137,"total_runtime_ms":565,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0111,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6373,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":524,"p95_latency_ms":641,"p99_latency_ms":687,"time_to_first_token_ms":154,"total_runtime_ms":582,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0264,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0914,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":477,"p95_latency_ms":582,"p99_latency_ms":625,"time_to_first_token_ms":146,"total_runtime_ms":529,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.016,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6116,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":486,"p95_latency_ms":595,"p99_latency_ms":638,"time_to_first_token_ms":158,"total_runtime_ms":540,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0194,"confidence_interval":0.9305,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.794,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":513,"p95_latency_ms":627,"p99_latency_ms":672,"time_to_first_token_ms":157,"total_runtime_ms":570,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":45,"gpu_utilization_percent":24,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.014,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3865,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":332,"p95_latency_ms":405,"p99_latency_ms":435,"time_to_first_token_ms":108,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.018,"confidence_interval":0.9419,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.113,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":329,"p95_latency_ms":402,"p99_latency_ms":431,"time_to_first_token_ms":100,"total_runtime_ms":366,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0258,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1249,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":353,"p95_latency_ms":431,"p99_latency_ms":463,"time_to_first_token_ms":114,"total_runtime_ms":392,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0249,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.422,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":275,"p95_latency_ms":336,"p99_latency_ms":360,"time_to_first_token_ms":103,"total_runtime_ms":305,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0134,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2127,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":298,"p95_latency_ms":364,"p99_latency_ms":390,"time_to_first_token_ms":100,"total_runtime_ms":331,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0277,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4283,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":255,"p95_latency_ms":312,"p99_latency_ms":334,"time_to_first_token_ms":119,"total_runtime_ms":283,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0206,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7547,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":345,"p95_latency_ms":421,"p99_latency_ms":452,"time_to_first_token_ms":114,"total_runtime_ms":383,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0254,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4026,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":343,"p95_latency_ms":419,"p99_latency_ms":450,"time_to_first_token_ms":95,"total_runtime_ms":381,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0147,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9995,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":313,"p95_latency_ms":383,"p99_latency_ms":411,"time_to_first_token_ms":89,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0184,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9164,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":349,"p95_latency_ms":427,"p99_latency_ms":458,"time_to_first_token_ms":93,"total_runtime_ms":388,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":35,"gpu_utilization_percent":24,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0232,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7247,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":804,"p95_latency_ms":983,"p99_latency_ms":1054,"time_to_first_token_ms":192,"total_runtime_ms":894,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0299,"confidence_interval":0.9441,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2157,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":808,"p95_latency_ms":987,"p99_latency_ms":1059,"time_to_first_token_ms":198,"total_runtime_ms":897,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0254,"confidence_interval":0.9399,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1918,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1015,"p99_latency_ms":1089,"time_to_first_token_ms":188,"total_runtime_ms":923,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0247,"confidence_interval":0.931,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0404,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":831,"p95_latency_ms":1016,"p99_latency_ms":1089,"time_to_first_token_ms":185,"total_runtime_ms":923,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0164,"confidence_interval":0.9456,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8736,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":792,"p95_latency_ms":968,"p99_latency_ms":1038,"time_to_first_token_ms":217,"total_runtime_ms":880,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0103,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.407,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":883,"p95_latency_ms":1080,"p99_latency_ms":1158,"time_to_first_token_ms":207,"total_runtime_ms":982,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0182,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5497,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":863,"p95_latency_ms":1054,"p99_latency_ms":1131,"time_to_first_token_ms":212,"total_runtime_ms":959,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0152,"confidence_interval":0.9404,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9976,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":858,"p95_latency_ms":1049,"p99_latency_ms":1125,"time_to_first_token_ms":213,"total_runtime_ms":953,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0138,"confidence_interval":0.9349,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1723,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1015,"p99_latency_ms":1088,"time_to_first_token_ms":196,"total_runtime_ms":922,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0269,"confidence_interval":0.9456,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7873,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":195,"total_runtime_ms":895,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":45,"gpu_utilization_percent":41,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0209,"confidence_interval":0.9489,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9433,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":631,"time_to_first_token_ms":164,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.024,"confidence_interval":0.9416,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9198,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":452,"p95_latency_ms":552,"p99_latency_ms":592,"time_to_first_token_ms":168,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0265,"confidence_interval":0.9442,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1397,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":495,"p95_latency_ms":605,"p99_latency_ms":649,"time_to_first_token_ms":152,"total_runtime_ms":550,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0248,"confidence_interval":0.9397,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7269,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":513,"p95_latency_ms":627,"p99_latency_ms":673,"time_to_first_token_ms":167,"total_runtime_ms":570,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0283,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9946,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":528,"p95_latency_ms":645,"p99_latency_ms":692,"time_to_first_token_ms":158,"total_runtime_ms":587,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0166,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3158,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":445,"p95_latency_ms":544,"p99_latency_ms":583,"time_to_first_token_ms":141,"total_runtime_ms":494,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0224,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0157,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":467,"p95_latency_ms":570,"p99_latency_ms":612,"time_to_first_token_ms":156,"total_runtime_ms":518,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0254,"confidence_interval":0.945,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.997,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":463,"p95_latency_ms":566,"p99_latency_ms":607,"time_to_first_token_ms":146,"total_runtime_ms":514,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.014,"confidence_interval":0.9417,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8521,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":453,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":140,"total_runtime_ms":503,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0688,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":499,"p95_latency_ms":610,"p99_latency_ms":655,"time_to_first_token_ms":146,"total_runtime_ms":555,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":35,"gpu_utilization_percent":41,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0196,"confidence_interval":0.9318,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3816,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":429,"p95_latency_ms":525,"p99_latency_ms":563,"time_to_first_token_ms":126,"total_runtime_ms":477,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9321,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6137,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":480,"p95_latency_ms":586,"p99_latency_ms":629,"time_to_first_token_ms":122,"total_runtime_ms":533,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0236,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5048,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":455,"p95_latency_ms":556,"p99_latency_ms":597,"time_to_first_token_ms":122,"total_runtime_ms":506,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0283,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0285,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":390,"p95_latency_ms":476,"p99_latency_ms":511,"time_to_first_token_ms":127,"total_runtime_ms":433,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0267,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8303,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":476,"p95_latency_ms":582,"p99_latency_ms":624,"time_to_first_token_ms":133,"total_runtime_ms":529,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3695,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":478,"p95_latency_ms":585,"p99_latency_ms":627,"time_to_first_token_ms":125,"total_runtime_ms":532,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.027,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.012,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":130,"total_runtime_ms":510,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.022,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4619,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":477,"p95_latency_ms":583,"p99_latency_ms":625,"time_to_first_token_ms":145,"total_runtime_ms":530,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0293,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2417,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":439,"p95_latency_ms":536,"p99_latency_ms":575,"time_to_first_token_ms":130,"total_runtime_ms":487,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0135,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.283,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":480,"p95_latency_ms":587,"p99_latency_ms":630,"time_to_first_token_ms":145,"total_runtime_ms":534,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0213,"confidence_interval":0.9473,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8634,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":219,"p95_latency_ms":268,"p99_latency_ms":287,"time_to_first_token_ms":87,"total_runtime_ms":243,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0281,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4161,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":270,"p95_latency_ms":331,"p99_latency_ms":355,"time_to_first_token_ms":104,"total_runtime_ms":300,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0298,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.893,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":244,"p95_latency_ms":299,"p99_latency_ms":321,"time_to_first_token_ms":96,"total_runtime_ms":272,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0155,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9643,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":282,"p95_latency_ms":345,"p99_latency_ms":370,"time_to_first_token_ms":81,"total_runtime_ms":313,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0199,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2934,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":300,"p95_latency_ms":367,"p99_latency_ms":394,"time_to_first_token_ms":105,"total_runtime_ms":334,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0141,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7418,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":357,"p99_latency_ms":383,"time_to_first_token_ms":85,"total_runtime_ms":325,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0233,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3148,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":244,"p95_latency_ms":298,"p99_latency_ms":320,"time_to_first_token_ms":88,"total_runtime_ms":271,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0291,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5698,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":219,"p95_latency_ms":267,"p99_latency_ms":286,"time_to_first_token_ms":94,"total_runtime_ms":243,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0285,"confidence_interval":0.9366,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0544,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":298,"p95_latency_ms":364,"p99_latency_ms":391,"time_to_first_token_ms":90,"total_runtime_ms":331,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4565,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":296,"p95_latency_ms":362,"p99_latency_ms":388,"time_to_first_token_ms":91,"total_runtime_ms":329,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0152,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2026,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":736,"p95_latency_ms":899,"p99_latency_ms":965,"time_to_first_token_ms":196,"total_runtime_ms":818,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0168,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7712,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":708,"p95_latency_ms":866,"p99_latency_ms":928,"time_to_first_token_ms":170,"total_runtime_ms":787,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0139,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2892,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":740,"p95_latency_ms":904,"p99_latency_ms":970,"time_to_first_token_ms":172,"total_runtime_ms":822,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0124,"confidence_interval":0.947,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6457,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":948,"time_to_first_token_ms":170,"total_runtime_ms":803,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0153,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0607,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":737,"p95_latency_ms":901,"p99_latency_ms":967,"time_to_first_token_ms":184,"total_runtime_ms":819,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0212,"confidence_interval":0.9468,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.81,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":704,"p95_latency_ms":860,"p99_latency_ms":923,"time_to_first_token_ms":170,"total_runtime_ms":782,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0277,"confidence_interval":0.9434,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6115,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":757,"p95_latency_ms":925,"p99_latency_ms":993,"time_to_first_token_ms":192,"total_runtime_ms":841,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0241,"confidence_interval":0.9482,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7656,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":700,"p95_latency_ms":856,"p99_latency_ms":918,"time_to_first_token_ms":165,"total_runtime_ms":778,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0248,"confidence_interval":0.9374,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3708,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":701,"p95_latency_ms":857,"p99_latency_ms":919,"time_to_first_token_ms":168,"total_runtime_ms":779,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0244,"confidence_interval":0.9365,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5963,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":755,"p95_latency_ms":923,"p99_latency_ms":990,"time_to_first_token_ms":188,"total_runtime_ms":839,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":39,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0176,"confidence_interval":0.9301,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3899,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":438,"p95_latency_ms":535,"p99_latency_ms":574,"time_to_first_token_ms":136,"total_runtime_ms":486,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6719,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":461,"p95_latency_ms":563,"p99_latency_ms":604,"time_to_first_token_ms":142,"total_runtime_ms":512,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0153,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0365,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":432,"p95_latency_ms":528,"p99_latency_ms":566,"time_to_first_token_ms":138,"total_runtime_ms":480,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0159,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2756,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":404,"p95_latency_ms":493,"p99_latency_ms":529,"time_to_first_token_ms":142,"total_runtime_ms":448,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0245,"confidence_interval":0.9402,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.089,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":475,"p99_latency_ms":510,"time_to_first_token_ms":136,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0247,"confidence_interval":0.9304,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9267,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":430,"p95_latency_ms":526,"p99_latency_ms":564,"time_to_first_token_ms":132,"total_runtime_ms":478,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0115,"confidence_interval":0.9341,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1621,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":427,"p95_latency_ms":522,"p99_latency_ms":560,"time_to_first_token_ms":133,"total_runtime_ms":475,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.021,"confidence_interval":0.9446,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8686,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":479,"time_to_first_token_ms":124,"total_runtime_ms":406,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0243,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7238,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":456,"p95_latency_ms":558,"p99_latency_ms":598,"time_to_first_token_ms":117,"total_runtime_ms":507,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0179,"confidence_interval":0.9392,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5271,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"classification_multiclass","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":435,"p95_latency_ms":531,"p99_latency_ms":570,"time_to_first_token_ms":135,"total_runtime_ms":483,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":29,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0244,"confidence_interval":0.9404,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9027,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1030,"p99_latency_ms":1105,"time_to_first_token_ms":219,"total_runtime_ms":936,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0168,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2742,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":209,"total_runtime_ms":889,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0254,"confidence_interval":0.9395,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7556,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":846,"p95_latency_ms":1034,"p99_latency_ms":1109,"time_to_first_token_ms":193,"total_runtime_ms":940,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0151,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3868,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":833,"p95_latency_ms":1018,"p99_latency_ms":1092,"time_to_first_token_ms":219,"total_runtime_ms":926,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0163,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9994,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":777,"p95_latency_ms":949,"p99_latency_ms":1018,"time_to_first_token_ms":215,"total_runtime_ms":863,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0297,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.638,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":850,"p95_latency_ms":1038,"p99_latency_ms":1114,"time_to_first_token_ms":220,"total_runtime_ms":944,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0254,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0133,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":759,"p95_latency_ms":928,"p99_latency_ms":995,"time_to_first_token_ms":217,"total_runtime_ms":843,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0221,"confidence_interval":0.9466,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9757,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":815,"p95_latency_ms":996,"p99_latency_ms":1069,"time_to_first_token_ms":189,"total_runtime_ms":906,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0249,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4579,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":801,"p95_latency_ms":979,"p99_latency_ms":1050,"time_to_first_token_ms":196,"total_runtime_ms":890,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0265,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8605,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":814,"p95_latency_ms":994,"p99_latency_ms":1067,"time_to_first_token_ms":215,"total_runtime_ms":904,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0249,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6967,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":400,"p95_latency_ms":489,"p99_latency_ms":524,"time_to_first_token_ms":143,"total_runtime_ms":444,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0265,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7414,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":407,"p95_latency_ms":498,"p99_latency_ms":534,"time_to_first_token_ms":149,"total_runtime_ms":452,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0117,"confidence_interval":0.9347,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8067,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":375,"p95_latency_ms":458,"p99_latency_ms":491,"time_to_first_token_ms":168,"total_runtime_ms":416,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.028,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2012,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":149,"total_runtime_ms":475,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0231,"confidence_interval":0.9479,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8588,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":421,"p95_latency_ms":515,"p99_latency_ms":552,"time_to_first_token_ms":162,"total_runtime_ms":468,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0295,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1135,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":397,"p95_latency_ms":486,"p99_latency_ms":521,"time_to_first_token_ms":150,"total_runtime_ms":442,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.021,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4411,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":476,"p99_latency_ms":510,"time_to_first_token_ms":163,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0287,"confidence_interval":0.9395,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0954,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":458,"p95_latency_ms":559,"p99_latency_ms":600,"time_to_first_token_ms":166,"total_runtime_ms":509,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0165,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0554,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":580,"p99_latency_ms":622,"time_to_first_token_ms":159,"total_runtime_ms":527,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0232,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.513,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":400,"p95_latency_ms":489,"p99_latency_ms":525,"time_to_first_token_ms":149,"total_runtime_ms":445,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0156,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1643,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1286,"p95_latency_ms":1572,"p99_latency_ms":1687,"time_to_first_token_ms":294,"total_runtime_ms":1429,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0119,"confidence_interval":0.931,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8812,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1357,"p95_latency_ms":1658,"p99_latency_ms":1779,"time_to_first_token_ms":301,"total_runtime_ms":1508,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0215,"confidence_interval":0.9357,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3684,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1342,"p95_latency_ms":1640,"p99_latency_ms":1759,"time_to_first_token_ms":292,"total_runtime_ms":1491,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0249,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6295,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1277,"p95_latency_ms":1561,"p99_latency_ms":1675,"time_to_first_token_ms":289,"total_runtime_ms":1419,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0294,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4134,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1351,"p95_latency_ms":1652,"p99_latency_ms":1772,"time_to_first_token_ms":286,"total_runtime_ms":1502,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0103,"confidence_interval":0.9476,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0636,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1375,"p95_latency_ms":1680,"p99_latency_ms":1802,"time_to_first_token_ms":286,"total_runtime_ms":1527,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0127,"confidence_interval":0.9394,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.174,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1365,"p95_latency_ms":1668,"p99_latency_ms":1789,"time_to_first_token_ms":311,"total_runtime_ms":1516,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0229,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2752,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1301,"p95_latency_ms":1590,"p99_latency_ms":1705,"time_to_first_token_ms":310,"total_runtime_ms":1445,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0195,"confidence_interval":0.9495,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0033,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1358,"p95_latency_ms":1660,"p99_latency_ms":1781,"time_to_first_token_ms":306,"total_runtime_ms":1509,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0158,"confidence_interval":0.934,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.7052,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1377,"p95_latency_ms":1684,"p99_latency_ms":1806,"time_to_first_token_ms":283,"total_runtime_ms":1531,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0216,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5814,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":639,"p95_latency_ms":781,"p99_latency_ms":838,"time_to_first_token_ms":248,"total_runtime_ms":710,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0198,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1539,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":744,"p95_latency_ms":909,"p99_latency_ms":975,"time_to_first_token_ms":266,"total_runtime_ms":827,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0113,"confidence_interval":0.935,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8189,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":724,"p95_latency_ms":885,"p99_latency_ms":949,"time_to_first_token_ms":247,"total_runtime_ms":805,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.026,"confidence_interval":0.9446,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7938,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":706,"p95_latency_ms":862,"p99_latency_ms":925,"time_to_first_token_ms":243,"total_runtime_ms":784,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0225,"confidence_interval":0.9406,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1204,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":637,"p95_latency_ms":779,"p99_latency_ms":835,"time_to_first_token_ms":247,"total_runtime_ms":708,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0155,"confidence_interval":0.9306,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2828,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":660,"p95_latency_ms":806,"p99_latency_ms":865,"time_to_first_token_ms":258,"total_runtime_ms":733,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0101,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2276,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":645,"p95_latency_ms":788,"p99_latency_ms":845,"time_to_first_token_ms":252,"total_runtime_ms":716,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0286,"confidence_interval":0.9343,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7809,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":664,"p95_latency_ms":811,"p99_latency_ms":870,"time_to_first_token_ms":266,"total_runtime_ms":738,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0224,"confidence_interval":0.9338,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.219,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":742,"p95_latency_ms":907,"p99_latency_ms":973,"time_to_first_token_ms":259,"total_runtime_ms":825,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0293,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6806,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":662,"p95_latency_ms":810,"p99_latency_ms":869,"time_to_first_token_ms":267,"total_runtime_ms":736,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0218,"confidence_interval":0.931,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8711,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":626,"p95_latency_ms":765,"p99_latency_ms":820,"time_to_first_token_ms":181,"total_runtime_ms":695,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0162,"confidence_interval":0.9369,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4167,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":657,"p95_latency_ms":802,"p99_latency_ms":861,"time_to_first_token_ms":176,"total_runtime_ms":730,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3918,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":600,"p95_latency_ms":733,"p99_latency_ms":786,"time_to_first_token_ms":179,"total_runtime_ms":666,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.015,"confidence_interval":0.95,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2273,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":743,"time_to_first_token_ms":190,"total_runtime_ms":630,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0124,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6379,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":773,"time_to_first_token_ms":165,"total_runtime_ms":655,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0132,"confidence_interval":0.9445,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1064,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":639,"p95_latency_ms":782,"p99_latency_ms":838,"time_to_first_token_ms":173,"total_runtime_ms":711,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0134,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8427,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":623,"p95_latency_ms":761,"p99_latency_ms":816,"time_to_first_token_ms":185,"total_runtime_ms":692,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.021,"confidence_interval":0.9305,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.114,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":568,"p95_latency_ms":694,"p99_latency_ms":745,"time_to_first_token_ms":157,"total_runtime_ms":631,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0159,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2684,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":620,"p95_latency_ms":758,"p99_latency_ms":813,"time_to_first_token_ms":168,"total_runtime_ms":689,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0179,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2687,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":640,"p95_latency_ms":782,"p99_latency_ms":839,"time_to_first_token_ms":164,"total_runtime_ms":711,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0235,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.207,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":320,"p95_latency_ms":392,"p99_latency_ms":420,"time_to_first_token_ms":113,"total_runtime_ms":356,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0174,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.174,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":299,"p95_latency_ms":366,"p99_latency_ms":392,"time_to_first_token_ms":125,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0162,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1018,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":306,"p95_latency_ms":374,"p99_latency_ms":401,"time_to_first_token_ms":117,"total_runtime_ms":340,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0183,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8878,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":290,"p95_latency_ms":354,"p99_latency_ms":380,"time_to_first_token_ms":114,"total_runtime_ms":322,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0238,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7942,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":385,"p95_latency_ms":471,"p99_latency_ms":505,"time_to_first_token_ms":128,"total_runtime_ms":428,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0181,"confidence_interval":0.9423,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9237,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":385,"p95_latency_ms":471,"p99_latency_ms":505,"time_to_first_token_ms":129,"total_runtime_ms":428,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0129,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4145,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":293,"p95_latency_ms":358,"p99_latency_ms":385,"time_to_first_token_ms":139,"total_runtime_ms":326,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8829,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":352,"p95_latency_ms":430,"p99_latency_ms":461,"time_to_first_token_ms":118,"total_runtime_ms":391,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0287,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.688,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":407,"time_to_first_token_ms":116,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0281,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8388,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":281,"p95_latency_ms":343,"p99_latency_ms":368,"time_to_first_token_ms":128,"total_runtime_ms":312,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0251,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3163,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":994,"p95_latency_ms":1215,"p99_latency_ms":1304,"time_to_first_token_ms":241,"total_runtime_ms":1105,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0129,"confidence_interval":0.9322,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9183,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1003,"p95_latency_ms":1226,"p99_latency_ms":1316,"time_to_first_token_ms":251,"total_runtime_ms":1115,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0252,"confidence_interval":0.9435,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4929,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1023,"p95_latency_ms":1250,"p99_latency_ms":1341,"time_to_first_token_ms":246,"total_runtime_ms":1137,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0156,"confidence_interval":0.938,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9109,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":990,"p95_latency_ms":1211,"p99_latency_ms":1299,"time_to_first_token_ms":231,"total_runtime_ms":1101,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0215,"confidence_interval":0.938,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.234,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":963,"p95_latency_ms":1176,"p99_latency_ms":1262,"time_to_first_token_ms":239,"total_runtime_ms":1070,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0109,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1927,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1018,"p95_latency_ms":1245,"p99_latency_ms":1335,"time_to_first_token_ms":249,"total_runtime_ms":1131,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0251,"confidence_interval":0.9405,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6256,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1007,"p95_latency_ms":1231,"p99_latency_ms":1321,"time_to_first_token_ms":255,"total_runtime_ms":1119,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0188,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4102,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1025,"p95_latency_ms":1253,"p99_latency_ms":1344,"time_to_first_token_ms":233,"total_runtime_ms":1139,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0269,"confidence_interval":0.95,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5159,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1026,"p95_latency_ms":1254,"p99_latency_ms":1345,"time_to_first_token_ms":251,"total_runtime_ms":1140,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.017,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8256,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":986,"p95_latency_ms":1205,"p99_latency_ms":1292,"time_to_first_token_ms":238,"total_runtime_ms":1095,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0208,"confidence_interval":0.9499,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9693,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":487,"p95_latency_ms":595,"p99_latency_ms":638,"time_to_first_token_ms":189,"total_runtime_ms":541,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0159,"confidence_interval":0.9411,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5787,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":574,"p95_latency_ms":701,"p99_latency_ms":752,"time_to_first_token_ms":185,"total_runtime_ms":638,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0187,"confidence_interval":0.9356,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6743,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":546,"p95_latency_ms":668,"p99_latency_ms":716,"time_to_first_token_ms":207,"total_runtime_ms":607,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0184,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.487,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":551,"p95_latency_ms":673,"p99_latency_ms":722,"time_to_first_token_ms":182,"total_runtime_ms":612,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0156,"confidence_interval":0.9421,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6227,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":564,"p95_latency_ms":689,"p99_latency_ms":740,"time_to_first_token_ms":180,"total_runtime_ms":627,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0124,"confidence_interval":0.9466,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7249,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":630,"time_to_first_token_ms":206,"total_runtime_ms":534,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0156,"confidence_interval":0.9383,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2181,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":555,"p95_latency_ms":678,"p99_latency_ms":727,"time_to_first_token_ms":203,"total_runtime_ms":616,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0242,"confidence_interval":0.9445,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0251,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":587,"p99_latency_ms":630,"time_to_first_token_ms":190,"total_runtime_ms":534,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0189,"confidence_interval":0.9382,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0031,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":583,"p95_latency_ms":713,"p99_latency_ms":765,"time_to_first_token_ms":194,"total_runtime_ms":648,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0251,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9598,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":524,"p95_latency_ms":640,"p99_latency_ms":687,"time_to_first_token_ms":183,"total_runtime_ms":582,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0164,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3935,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":446,"p95_latency_ms":546,"p99_latency_ms":585,"time_to_first_token_ms":134,"total_runtime_ms":496,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0292,"confidence_interval":0.9442,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4366,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":465,"p95_latency_ms":569,"p99_latency_ms":610,"time_to_first_token_ms":135,"total_runtime_ms":517,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0214,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.146,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":431,"p95_latency_ms":527,"p99_latency_ms":565,"time_to_first_token_ms":137,"total_runtime_ms":479,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0109,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8193,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":504,"p95_latency_ms":616,"p99_latency_ms":661,"time_to_first_token_ms":133,"total_runtime_ms":560,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.012,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4481,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":492,"p95_latency_ms":602,"p99_latency_ms":645,"time_to_first_token_ms":146,"total_runtime_ms":547,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0217,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6394,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":493,"p95_latency_ms":602,"p99_latency_ms":646,"time_to_first_token_ms":150,"total_runtime_ms":547,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0155,"confidence_interval":0.9468,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8421,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":447,"p95_latency_ms":546,"p99_latency_ms":586,"time_to_first_token_ms":161,"total_runtime_ms":497,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0223,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3229,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":491,"p95_latency_ms":600,"p99_latency_ms":644,"time_to_first_token_ms":157,"total_runtime_ms":546,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0127,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0212,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":167,"total_runtime_ms":476,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0217,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5278,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":595,"time_to_first_token_ms":142,"total_runtime_ms":504,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.023,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5338,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":315,"p95_latency_ms":384,"p99_latency_ms":412,"time_to_first_token_ms":111,"total_runtime_ms":350,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0208,"confidence_interval":0.9428,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2012,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":295,"p95_latency_ms":360,"p99_latency_ms":387,"time_to_first_token_ms":108,"total_runtime_ms":328,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0109,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1829,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":299,"p95_latency_ms":365,"p99_latency_ms":392,"time_to_first_token_ms":92,"total_runtime_ms":332,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0157,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6481,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":272,"p95_latency_ms":333,"p99_latency_ms":357,"time_to_first_token_ms":108,"total_runtime_ms":302,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0187,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3176,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":314,"p95_latency_ms":384,"p99_latency_ms":412,"time_to_first_token_ms":91,"total_runtime_ms":349,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0154,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8064,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":249,"p95_latency_ms":305,"p99_latency_ms":327,"time_to_first_token_ms":105,"total_runtime_ms":277,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0241,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3364,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":259,"p95_latency_ms":316,"p99_latency_ms":339,"time_to_first_token_ms":115,"total_runtime_ms":287,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8872,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":357,"p99_latency_ms":382,"time_to_first_token_ms":97,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0187,"confidence_interval":0.9429,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9843,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":212,"p95_latency_ms":259,"p99_latency_ms":278,"time_to_first_token_ms":117,"total_runtime_ms":235,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0229,"confidence_interval":0.9349,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8703,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":261,"p95_latency_ms":320,"p99_latency_ms":343,"time_to_first_token_ms":118,"total_runtime_ms":290,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0289,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1322,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":816,"p95_latency_ms":997,"p99_latency_ms":1070,"time_to_first_token_ms":201,"total_runtime_ms":906,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0295,"confidence_interval":0.9457,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4142,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":741,"p95_latency_ms":906,"p99_latency_ms":971,"time_to_first_token_ms":213,"total_runtime_ms":823,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0159,"confidence_interval":0.9411,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2125,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":767,"p95_latency_ms":938,"p99_latency_ms":1006,"time_to_first_token_ms":187,"total_runtime_ms":852,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.025,"confidence_interval":0.9391,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1215,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":722,"p95_latency_ms":883,"p99_latency_ms":947,"time_to_first_token_ms":211,"total_runtime_ms":802,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0195,"confidence_interval":0.9416,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6372,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":754,"p95_latency_ms":922,"p99_latency_ms":989,"time_to_first_token_ms":203,"total_runtime_ms":838,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0105,"confidence_interval":0.9457,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6691,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":750,"p95_latency_ms":916,"p99_latency_ms":983,"time_to_first_token_ms":193,"total_runtime_ms":833,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.024,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1507,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":766,"p95_latency_ms":937,"p99_latency_ms":1005,"time_to_first_token_ms":185,"total_runtime_ms":852,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0187,"confidence_interval":0.9358,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4881,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":799,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":183,"total_runtime_ms":887,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0287,"confidence_interval":0.9462,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5529,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":740,"p95_latency_ms":905,"p99_latency_ms":970,"time_to_first_token_ms":198,"total_runtime_ms":822,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0186,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8666,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":815,"p95_latency_ms":996,"p99_latency_ms":1069,"time_to_first_token_ms":207,"total_runtime_ms":906,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.03,"confidence_interval":0.9364,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3149,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":465,"p95_latency_ms":568,"p99_latency_ms":609,"time_to_first_token_ms":139,"total_runtime_ms":516,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0269,"confidence_interval":0.9372,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7598,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":591,"time_to_first_token_ms":161,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.023,"confidence_interval":0.9392,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4015,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":478,"time_to_first_token_ms":142,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0292,"confidence_interval":0.9462,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5566,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":374,"p95_latency_ms":457,"p99_latency_ms":490,"time_to_first_token_ms":164,"total_runtime_ms":415,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0285,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5554,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":460,"p95_latency_ms":562,"p99_latency_ms":603,"time_to_first_token_ms":164,"total_runtime_ms":511,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0259,"confidence_interval":0.9334,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7848,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":433,"p95_latency_ms":529,"p99_latency_ms":568,"time_to_first_token_ms":153,"total_runtime_ms":481,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0264,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2915,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":166,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0288,"confidence_interval":0.9311,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9135,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":393,"p95_latency_ms":481,"p99_latency_ms":516,"time_to_first_token_ms":153,"total_runtime_ms":437,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0143,"confidence_interval":0.9344,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4734,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":406,"p95_latency_ms":496,"p99_latency_ms":532,"time_to_first_token_ms":141,"total_runtime_ms":451,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0118,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0113,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":458,"p95_latency_ms":559,"p99_latency_ms":600,"time_to_first_token_ms":157,"total_runtime_ms":508,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0153,"confidence_interval":0.9377,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8759,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":137,"total_runtime_ms":453,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4755,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":351,"p95_latency_ms":429,"p99_latency_ms":460,"time_to_first_token_ms":152,"total_runtime_ms":390,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0143,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5355,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":437,"p95_latency_ms":534,"p99_latency_ms":572,"time_to_first_token_ms":150,"total_runtime_ms":485,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0233,"confidence_interval":0.9472,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3849,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":442,"p95_latency_ms":541,"p99_latency_ms":580,"time_to_first_token_ms":154,"total_runtime_ms":492,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0105,"confidence_interval":0.9361,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0313,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":453,"p95_latency_ms":554,"p99_latency_ms":595,"time_to_first_token_ms":123,"total_runtime_ms":504,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0173,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6865,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":508,"time_to_first_token_ms":145,"total_runtime_ms":430,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.015,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3958,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":384,"p95_latency_ms":470,"p99_latency_ms":504,"time_to_first_token_ms":153,"total_runtime_ms":427,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0158,"confidence_interval":0.9408,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3838,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":401,"p95_latency_ms":490,"p99_latency_ms":526,"time_to_first_token_ms":140,"total_runtime_ms":445,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0255,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7257,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":146,"total_runtime_ms":454,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0156,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1526,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":447,"p95_latency_ms":547,"p99_latency_ms":587,"time_to_first_token_ms":148,"total_runtime_ms":497,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0164,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7898,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":215,"p95_latency_ms":263,"p99_latency_ms":283,"time_to_first_token_ms":81,"total_runtime_ms":239,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0102,"confidence_interval":0.938,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1854,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":261,"p95_latency_ms":319,"p99_latency_ms":343,"time_to_first_token_ms":98,"total_runtime_ms":290,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0267,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5452,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":281,"p95_latency_ms":343,"p99_latency_ms":368,"time_to_first_token_ms":98,"total_runtime_ms":312,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0247,"confidence_interval":0.9434,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1824,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":277,"p95_latency_ms":338,"p99_latency_ms":363,"time_to_first_token_ms":107,"total_runtime_ms":307,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0281,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2702,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":262,"p95_latency_ms":321,"p99_latency_ms":344,"time_to_first_token_ms":79,"total_runtime_ms":292,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0237,"confidence_interval":0.9347,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7305,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":189,"p95_latency_ms":231,"p99_latency_ms":248,"time_to_first_token_ms":107,"total_runtime_ms":210,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0116,"confidence_interval":0.9407,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3111,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":233,"p95_latency_ms":285,"p99_latency_ms":305,"time_to_first_token_ms":78,"total_runtime_ms":259,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0212,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2112,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":179,"p95_latency_ms":219,"p99_latency_ms":235,"time_to_first_token_ms":91,"total_runtime_ms":199,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.029,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2717,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":198,"p95_latency_ms":242,"p99_latency_ms":259,"time_to_first_token_ms":104,"total_runtime_ms":220,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0113,"confidence_interval":0.9453,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3752,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":228,"p95_latency_ms":279,"p99_latency_ms":299,"time_to_first_token_ms":106,"total_runtime_ms":253,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0195,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7394,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":691,"p95_latency_ms":844,"p99_latency_ms":906,"time_to_first_token_ms":189,"total_runtime_ms":768,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0144,"confidence_interval":0.9322,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1071,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":693,"p95_latency_ms":847,"p99_latency_ms":909,"time_to_first_token_ms":172,"total_runtime_ms":770,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0229,"confidence_interval":0.941,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9995,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":672,"p95_latency_ms":822,"p99_latency_ms":882,"time_to_first_token_ms":164,"total_runtime_ms":747,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0214,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8741,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":634,"p95_latency_ms":775,"p99_latency_ms":831,"time_to_first_token_ms":172,"total_runtime_ms":704,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0146,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2745,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":646,"p95_latency_ms":789,"p99_latency_ms":847,"time_to_first_token_ms":176,"total_runtime_ms":718,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0251,"confidence_interval":0.9371,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6433,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":700,"p95_latency_ms":855,"p99_latency_ms":918,"time_to_first_token_ms":191,"total_runtime_ms":778,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0238,"confidence_interval":0.943,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2153,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":670,"p95_latency_ms":819,"p99_latency_ms":878,"time_to_first_token_ms":179,"total_runtime_ms":744,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0161,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3491,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":678,"p95_latency_ms":829,"p99_latency_ms":889,"time_to_first_token_ms":164,"total_runtime_ms":753,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0191,"confidence_interval":0.9466,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6004,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":703,"p95_latency_ms":859,"p99_latency_ms":922,"time_to_first_token_ms":192,"total_runtime_ms":781,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.026,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6183,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":659,"p95_latency_ms":806,"p99_latency_ms":864,"time_to_first_token_ms":169,"total_runtime_ms":732,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0281,"confidence_interval":0.9338,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6332,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":366,"p95_latency_ms":448,"p99_latency_ms":480,"time_to_first_token_ms":126,"total_runtime_ms":407,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9413,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2684,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":397,"p95_latency_ms":486,"p99_latency_ms":521,"time_to_first_token_ms":148,"total_runtime_ms":442,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0236,"confidence_interval":0.942,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7352,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":312,"p95_latency_ms":381,"p99_latency_ms":409,"time_to_first_token_ms":147,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0266,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3221,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":403,"p95_latency_ms":493,"p99_latency_ms":529,"time_to_first_token_ms":126,"total_runtime_ms":448,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0221,"confidence_interval":0.9382,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8652,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":302,"p95_latency_ms":369,"p99_latency_ms":396,"time_to_first_token_ms":118,"total_runtime_ms":335,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0274,"confidence_interval":0.9345,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8908,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":384,"p95_latency_ms":469,"p99_latency_ms":503,"time_to_first_token_ms":146,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9366,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6144,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":336,"p95_latency_ms":410,"p99_latency_ms":440,"time_to_first_token_ms":141,"total_runtime_ms":373,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0202,"confidence_interval":0.9404,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0652,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":304,"p95_latency_ms":371,"p99_latency_ms":398,"time_to_first_token_ms":135,"total_runtime_ms":337,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0258,"confidence_interval":0.9329,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9671,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":347,"p95_latency_ms":424,"p99_latency_ms":455,"time_to_first_token_ms":139,"total_runtime_ms":385,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.013,"confidence_interval":0.9327,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8679,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_entities|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_entities","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":304,"p95_latency_ms":372,"p99_latency_ms":399,"time_to_first_token_ms":139,"total_runtime_ms":338,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0246,"confidence_interval":0.9448,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3493,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":939,"p95_latency_ms":1147,"p99_latency_ms":1231,"time_to_first_token_ms":203,"total_runtime_ms":1043,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0193,"confidence_interval":0.93,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2629,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":929,"p95_latency_ms":1135,"p99_latency_ms":1218,"time_to_first_token_ms":189,"total_runtime_ms":1032,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0194,"confidence_interval":0.941,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5564,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":912,"p95_latency_ms":1115,"p99_latency_ms":1196,"time_to_first_token_ms":214,"total_runtime_ms":1014,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0129,"confidence_interval":0.9329,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8162,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":922,"p95_latency_ms":1127,"p99_latency_ms":1209,"time_to_first_token_ms":211,"total_runtime_ms":1025,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0251,"confidence_interval":0.9364,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0158,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":927,"p95_latency_ms":1133,"p99_latency_ms":1216,"time_to_first_token_ms":200,"total_runtime_ms":1030,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0209,"confidence_interval":0.9454,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6266,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":893,"p95_latency_ms":1092,"p99_latency_ms":1171,"time_to_first_token_ms":195,"total_runtime_ms":992,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0119,"confidence_interval":0.9363,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.529,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":905,"p95_latency_ms":1106,"p99_latency_ms":1186,"time_to_first_token_ms":206,"total_runtime_ms":1005,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0106,"confidence_interval":0.9495,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4109,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":903,"p95_latency_ms":1103,"p99_latency_ms":1183,"time_to_first_token_ms":214,"total_runtime_ms":1003,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0155,"confidence_interval":0.9399,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4474,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":946,"p95_latency_ms":1156,"p99_latency_ms":1240,"time_to_first_token_ms":194,"total_runtime_ms":1051,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0142,"confidence_interval":0.9492,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3194,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":936,"p95_latency_ms":1145,"p99_latency_ms":1228,"time_to_first_token_ms":204,"total_runtime_ms":1041,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0298,"confidence_interval":0.9407,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7264,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":548,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":169,"total_runtime_ms":608,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.9347,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0154,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":569,"p95_latency_ms":695,"p99_latency_ms":746,"time_to_first_token_ms":151,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0129,"confidence_interval":0.9449,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0025,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":529,"p95_latency_ms":646,"p99_latency_ms":693,"time_to_first_token_ms":176,"total_runtime_ms":588,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0218,"confidence_interval":0.9326,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4758,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":521,"p95_latency_ms":637,"p99_latency_ms":684,"time_to_first_token_ms":153,"total_runtime_ms":579,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.028,"confidence_interval":0.9332,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1955,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":524,"p95_latency_ms":641,"p99_latency_ms":687,"time_to_first_token_ms":163,"total_runtime_ms":582,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0208,"confidence_interval":0.9489,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5168,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":602,"p95_latency_ms":736,"p99_latency_ms":789,"time_to_first_token_ms":176,"total_runtime_ms":669,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0254,"confidence_interval":0.9407,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7922,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":538,"p95_latency_ms":657,"p99_latency_ms":705,"time_to_first_token_ms":178,"total_runtime_ms":597,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.023,"confidence_interval":0.9335,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1449,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":588,"p95_latency_ms":718,"p99_latency_ms":770,"time_to_first_token_ms":178,"total_runtime_ms":653,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0214,"confidence_interval":0.9326,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4235,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":772,"time_to_first_token_ms":158,"total_runtime_ms":654,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0119,"confidence_interval":0.935,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1482,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":581,"p95_latency_ms":710,"p99_latency_ms":762,"time_to_first_token_ms":157,"total_runtime_ms":646,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0245,"confidence_interval":0.9406,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4055,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1587,"p95_latency_ms":1940,"p99_latency_ms":2081,"time_to_first_token_ms":297,"total_runtime_ms":1763,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0262,"confidence_interval":0.9393,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7482,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1572,"p95_latency_ms":1921,"p99_latency_ms":2061,"time_to_first_token_ms":302,"total_runtime_ms":1746,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0257,"confidence_interval":0.932,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7957,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1518,"p95_latency_ms":1855,"p99_latency_ms":1990,"time_to_first_token_ms":299,"total_runtime_ms":1687,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0146,"confidence_interval":0.9471,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0734,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1535,"p95_latency_ms":1876,"p99_latency_ms":2012,"time_to_first_token_ms":304,"total_runtime_ms":1705,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0166,"confidence_interval":0.9431,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.312,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1557,"p95_latency_ms":1903,"p99_latency_ms":2042,"time_to_first_token_ms":281,"total_runtime_ms":1730,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0145,"confidence_interval":0.9462,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1313,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1578,"p95_latency_ms":1929,"p99_latency_ms":2069,"time_to_first_token_ms":299,"total_runtime_ms":1753,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0135,"confidence_interval":0.9477,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2674,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1530,"p95_latency_ms":1870,"p99_latency_ms":2006,"time_to_first_token_ms":278,"total_runtime_ms":1700,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0224,"confidence_interval":0.9409,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1894,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1506,"p95_latency_ms":1840,"p99_latency_ms":1974,"time_to_first_token_ms":282,"total_runtime_ms":1673,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0101,"confidence_interval":0.9495,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6286,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1496,"p95_latency_ms":1829,"p99_latency_ms":1962,"time_to_first_token_ms":303,"total_runtime_ms":1663,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0182,"confidence_interval":0.9439,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1327,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1525,"p95_latency_ms":1864,"p99_latency_ms":2000,"time_to_first_token_ms":280,"total_runtime_ms":1695,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0141,"confidence_interval":0.9471,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3939,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":947,"p95_latency_ms":1158,"p99_latency_ms":1242,"time_to_first_token_ms":256,"total_runtime_ms":1052,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0163,"confidence_interval":0.9367,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.7174,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":921,"p95_latency_ms":1126,"p99_latency_ms":1208,"time_to_first_token_ms":237,"total_runtime_ms":1024,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0123,"confidence_interval":0.9372,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5039,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":940,"p95_latency_ms":1149,"p99_latency_ms":1233,"time_to_first_token_ms":243,"total_runtime_ms":1045,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0103,"confidence_interval":0.9475,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4087,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":914,"p95_latency_ms":1117,"p99_latency_ms":1199,"time_to_first_token_ms":258,"total_runtime_ms":1016,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0253,"confidence_interval":0.9358,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0708,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":866,"p95_latency_ms":1059,"p99_latency_ms":1136,"time_to_first_token_ms":265,"total_runtime_ms":963,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0232,"confidence_interval":0.9385,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6601,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":962,"p95_latency_ms":1176,"p99_latency_ms":1261,"time_to_first_token_ms":245,"total_runtime_ms":1069,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0192,"confidence_interval":0.9341,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9107,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":927,"p95_latency_ms":1133,"p99_latency_ms":1215,"time_to_first_token_ms":255,"total_runtime_ms":1030,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0104,"confidence_interval":0.9328,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8981,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":907,"p95_latency_ms":1108,"p99_latency_ms":1189,"time_to_first_token_ms":237,"total_runtime_ms":1008,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0299,"confidence_interval":0.9403,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1342,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":885,"p95_latency_ms":1082,"p99_latency_ms":1160,"time_to_first_token_ms":239,"total_runtime_ms":983,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0134,"confidence_interval":0.9435,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2755,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":896,"p95_latency_ms":1096,"p99_latency_ms":1175,"time_to_first_token_ms":261,"total_runtime_ms":996,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0284,"confidence_interval":0.9489,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0074,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":668,"p95_latency_ms":816,"p99_latency_ms":876,"time_to_first_token_ms":180,"total_runtime_ms":742,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0113,"confidence_interval":0.9465,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9361,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":705,"p95_latency_ms":862,"p99_latency_ms":925,"time_to_first_token_ms":171,"total_runtime_ms":784,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0218,"confidence_interval":0.9445,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1211,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":685,"p95_latency_ms":838,"p99_latency_ms":898,"time_to_first_token_ms":166,"total_runtime_ms":761,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0241,"confidence_interval":0.9455,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.055,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":735,"p95_latency_ms":898,"p99_latency_ms":963,"time_to_first_token_ms":168,"total_runtime_ms":816,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.014,"confidence_interval":0.9408,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6512,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":746,"p95_latency_ms":912,"p99_latency_ms":979,"time_to_first_token_ms":158,"total_runtime_ms":829,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0172,"confidence_interval":0.9359,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6695,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":719,"p95_latency_ms":879,"p99_latency_ms":942,"time_to_first_token_ms":186,"total_runtime_ms":799,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.012,"confidence_interval":0.9386,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3622,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":670,"p95_latency_ms":819,"p99_latency_ms":878,"time_to_first_token_ms":172,"total_runtime_ms":744,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0259,"confidence_interval":0.933,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0306,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":734,"p95_latency_ms":897,"p99_latency_ms":963,"time_to_first_token_ms":175,"total_runtime_ms":816,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.9395,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.126,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":757,"p95_latency_ms":926,"p99_latency_ms":993,"time_to_first_token_ms":188,"total_runtime_ms":841,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0204,"confidence_interval":0.9454,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7926,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":718,"p95_latency_ms":878,"p99_latency_ms":942,"time_to_first_token_ms":187,"total_runtime_ms":798,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.9415,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7349,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":395,"p95_latency_ms":483,"p99_latency_ms":518,"time_to_first_token_ms":116,"total_runtime_ms":439,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0175,"confidence_interval":0.9472,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8226,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":422,"p95_latency_ms":516,"p99_latency_ms":554,"time_to_first_token_ms":128,"total_runtime_ms":469,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.029,"confidence_interval":0.9308,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4753,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":444,"p95_latency_ms":543,"p99_latency_ms":583,"time_to_first_token_ms":143,"total_runtime_ms":494,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0124,"confidence_interval":0.9442,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5781,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":417,"p95_latency_ms":510,"p99_latency_ms":547,"time_to_first_token_ms":137,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0199,"confidence_interval":0.9447,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4763,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":460,"p95_latency_ms":562,"p99_latency_ms":602,"time_to_first_token_ms":129,"total_runtime_ms":511,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0126,"confidence_interval":0.9383,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.922,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":399,"p95_latency_ms":488,"p99_latency_ms":524,"time_to_first_token_ms":131,"total_runtime_ms":444,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0203,"confidence_interval":0.9418,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2114,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":120,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0242,"confidence_interval":0.9427,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2735,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":114,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0235,"confidence_interval":0.9337,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3745,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":434,"p95_latency_ms":530,"p99_latency_ms":569,"time_to_first_token_ms":128,"total_runtime_ms":482,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0242,"confidence_interval":0.9328,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9517,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":381,"p95_latency_ms":465,"p99_latency_ms":499,"time_to_first_token_ms":136,"total_runtime_ms":423,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0174,"confidence_interval":0.9352,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9694,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1213,"p95_latency_ms":1482,"p99_latency_ms":1590,"time_to_first_token_ms":242,"total_runtime_ms":1348,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0134,"confidence_interval":0.9482,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1285,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1198,"p95_latency_ms":1464,"p99_latency_ms":1571,"time_to_first_token_ms":229,"total_runtime_ms":1331,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9409,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0818,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1149,"p95_latency_ms":1404,"p99_latency_ms":1506,"time_to_first_token_ms":241,"total_runtime_ms":1276,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0217,"confidence_interval":0.9498,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5176,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1172,"p95_latency_ms":1433,"p99_latency_ms":1537,"time_to_first_token_ms":229,"total_runtime_ms":1303,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0113,"confidence_interval":0.9413,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":6.797,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1169,"p95_latency_ms":1429,"p99_latency_ms":1533,"time_to_first_token_ms":243,"total_runtime_ms":1299,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0162,"confidence_interval":0.9459,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6244,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1169,"p95_latency_ms":1428,"p99_latency_ms":1532,"time_to_first_token_ms":257,"total_runtime_ms":1299,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0105,"confidence_interval":0.936,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0909,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1159,"p95_latency_ms":1416,"p99_latency_ms":1519,"time_to_first_token_ms":248,"total_runtime_ms":1287,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9453,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9113,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1177,"p95_latency_ms":1439,"p99_latency_ms":1543,"time_to_first_token_ms":245,"total_runtime_ms":1308,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0129,"confidence_interval":0.9411,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3405,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1227,"p95_latency_ms":1500,"p99_latency_ms":1609,"time_to_first_token_ms":255,"total_runtime_ms":1364,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0207,"confidence_interval":0.93,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6469,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1124,"p95_latency_ms":1374,"p99_latency_ms":1473,"time_to_first_token_ms":239,"total_runtime_ms":1249,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0256,"confidence_interval":0.9463,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5849,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":717,"p95_latency_ms":877,"p99_latency_ms":941,"time_to_first_token_ms":198,"total_runtime_ms":797,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0143,"confidence_interval":0.9497,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6696,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":693,"p95_latency_ms":847,"p99_latency_ms":908,"time_to_first_token_ms":180,"total_runtime_ms":770,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.011,"confidence_interval":0.9376,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7522,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":705,"p95_latency_ms":861,"p99_latency_ms":924,"time_to_first_token_ms":202,"total_runtime_ms":783,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0148,"confidence_interval":0.9475,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1693,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":726,"p95_latency_ms":887,"p99_latency_ms":951,"time_to_first_token_ms":196,"total_runtime_ms":806,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0103,"confidence_interval":0.9473,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1103,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":663,"p95_latency_ms":811,"p99_latency_ms":870,"time_to_first_token_ms":202,"total_runtime_ms":737,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0102,"confidence_interval":0.9468,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8281,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":660,"p95_latency_ms":807,"p99_latency_ms":866,"time_to_first_token_ms":196,"total_runtime_ms":734,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0215,"confidence_interval":0.9318,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2964,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":669,"p95_latency_ms":817,"p99_latency_ms":877,"time_to_first_token_ms":196,"total_runtime_ms":743,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0295,"confidence_interval":0.948,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.549,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":735,"p95_latency_ms":899,"p99_latency_ms":964,"time_to_first_token_ms":185,"total_runtime_ms":817,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0249,"confidence_interval":0.945,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7242,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":670,"p95_latency_ms":819,"p99_latency_ms":878,"time_to_first_token_ms":190,"total_runtime_ms":744,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.011,"confidence_interval":0.9382,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8618,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":649,"p95_latency_ms":794,"p99_latency_ms":851,"time_to_first_token_ms":202,"total_runtime_ms":721,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0103,"confidence_interval":0.9463,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4121,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":671,"time_to_first_token_ms":159,"total_runtime_ms":569,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0286,"confidence_interval":0.9314,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5474,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":514,"p95_latency_ms":628,"p99_latency_ms":673,"time_to_first_token_ms":153,"total_runtime_ms":571,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0236,"confidence_interval":0.93,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1509,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":134,"total_runtime_ms":586,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0283,"confidence_interval":0.9456,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9616,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":586,"p95_latency_ms":717,"p99_latency_ms":769,"time_to_first_token_ms":165,"total_runtime_ms":652,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0103,"confidence_interval":0.9467,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8408,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":564,"p95_latency_ms":689,"p99_latency_ms":739,"time_to_first_token_ms":137,"total_runtime_ms":627,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0152,"confidence_interval":0.9451,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6711,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":562,"p95_latency_ms":686,"p99_latency_ms":736,"time_to_first_token_ms":142,"total_runtime_ms":624,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0128,"confidence_interval":0.9422,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0633,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":553,"p95_latency_ms":676,"p99_latency_ms":725,"time_to_first_token_ms":157,"total_runtime_ms":615,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0196,"confidence_interval":0.9412,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3135,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":668,"time_to_first_token_ms":146,"total_runtime_ms":566,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0112,"confidence_interval":0.945,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8517,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":560,"p95_latency_ms":684,"p99_latency_ms":734,"time_to_first_token_ms":147,"total_runtime_ms":622,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0243,"confidence_interval":0.9433,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4178,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":584,"p95_latency_ms":714,"p99_latency_ms":766,"time_to_first_token_ms":161,"total_runtime_ms":649,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0234,"confidence_interval":0.9369,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8564,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":389,"p95_latency_ms":476,"p99_latency_ms":510,"time_to_first_token_ms":113,"total_runtime_ms":433,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0108,"confidence_interval":0.9355,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7635,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":286,"p95_latency_ms":350,"p99_latency_ms":375,"time_to_first_token_ms":95,"total_runtime_ms":318,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0232,"confidence_interval":0.9407,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2143,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":375,"p95_latency_ms":458,"p99_latency_ms":492,"time_to_first_token_ms":105,"total_runtime_ms":417,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0119,"confidence_interval":0.9306,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1037,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":483,"time_to_first_token_ms":114,"total_runtime_ms":409,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9383,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.532,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":331,"p95_latency_ms":404,"p99_latency_ms":434,"time_to_first_token_ms":120,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0264,"confidence_interval":0.938,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5228,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":323,"p95_latency_ms":394,"p99_latency_ms":423,"time_to_first_token_ms":115,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0275,"confidence_interval":0.9353,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1172,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":307,"p95_latency_ms":376,"p99_latency_ms":403,"time_to_first_token_ms":102,"total_runtime_ms":342,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.026,"confidence_interval":0.9379,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.596,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":302,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":109,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0176,"confidence_interval":0.9476,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2916,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":507,"time_to_first_token_ms":115,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.03,"confidence_interval":0.945,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4405,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":379,"p95_latency_ms":463,"p99_latency_ms":496,"time_to_first_token_ms":98,"total_runtime_ms":421,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0207,"confidence_interval":0.9383,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.041,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":883,"p95_latency_ms":1079,"p99_latency_ms":1157,"time_to_first_token_ms":209,"total_runtime_ms":981,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0244,"confidence_interval":0.9408,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8501,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":922,"p95_latency_ms":1127,"p99_latency_ms":1208,"time_to_first_token_ms":204,"total_runtime_ms":1024,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.9327,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3759,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":844,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":201,"total_runtime_ms":937,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0271,"confidence_interval":0.9356,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.13,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":905,"p95_latency_ms":1106,"p99_latency_ms":1186,"time_to_first_token_ms":216,"total_runtime_ms":1005,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0117,"confidence_interval":0.9325,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8265,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":931,"p95_latency_ms":1138,"p99_latency_ms":1221,"time_to_first_token_ms":206,"total_runtime_ms":1035,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0193,"confidence_interval":0.9345,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9194,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":889,"p95_latency_ms":1086,"p99_latency_ms":1165,"time_to_first_token_ms":201,"total_runtime_ms":987,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0136,"confidence_interval":0.9391,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3115,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":862,"p95_latency_ms":1053,"p99_latency_ms":1130,"time_to_first_token_ms":195,"total_runtime_ms":957,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0174,"confidence_interval":0.9422,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5108,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":865,"p95_latency_ms":1057,"p99_latency_ms":1134,"time_to_first_token_ms":204,"total_runtime_ms":961,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0247,"confidence_interval":0.9337,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0236,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":890,"p95_latency_ms":1087,"p99_latency_ms":1166,"time_to_first_token_ms":191,"total_runtime_ms":988,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.9458,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2997,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":879,"p95_latency_ms":1074,"p99_latency_ms":1152,"time_to_first_token_ms":213,"total_runtime_ms":976,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0108,"confidence_interval":0.9413,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5267,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":672,"time_to_first_token_ms":152,"total_runtime_ms":569,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.937,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":6.919,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":519,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":166,"total_runtime_ms":577,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0116,"confidence_interval":0.9338,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4858,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":522,"p95_latency_ms":638,"p99_latency_ms":685,"time_to_first_token_ms":149,"total_runtime_ms":580,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0299,"confidence_interval":0.9333,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9059,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":568,"p95_latency_ms":694,"p99_latency_ms":744,"time_to_first_token_ms":169,"total_runtime_ms":631,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.9442,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":6.752,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":513,"p95_latency_ms":627,"p99_latency_ms":673,"time_to_first_token_ms":151,"total_runtime_ms":570,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0235,"confidence_interval":0.9449,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2126,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":547,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":151,"total_runtime_ms":608,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0137,"confidence_interval":0.9381,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":6.742,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":555,"p95_latency_ms":679,"p99_latency_ms":728,"time_to_first_token_ms":144,"total_runtime_ms":617,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.027,"confidence_interval":0.9495,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2541,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":582,"p95_latency_ms":711,"p99_latency_ms":763,"time_to_first_token_ms":153,"total_runtime_ms":647,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0125,"confidence_interval":0.9357,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4521,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":163,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0224,"confidence_interval":0.9371,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7419,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":568,"p95_latency_ms":695,"p99_latency_ms":745,"time_to_first_token_ms":141,"total_runtime_ms":632,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0157,"confidence_interval":0.9364,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9393,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":512,"p95_latency_ms":625,"p99_latency_ms":671,"time_to_first_token_ms":134,"total_runtime_ms":569,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0273,"confidence_interval":0.9393,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9384,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":457,"p95_latency_ms":558,"p99_latency_ms":599,"time_to_first_token_ms":145,"total_runtime_ms":507,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0229,"confidence_interval":0.9356,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8234,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":443,"p95_latency_ms":541,"p99_latency_ms":580,"time_to_first_token_ms":128,"total_runtime_ms":492,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0106,"confidence_interval":0.9334,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0207,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":427,"p95_latency_ms":522,"p99_latency_ms":560,"time_to_first_token_ms":148,"total_runtime_ms":474,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0247,"confidence_interval":0.9344,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3027,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":474,"p95_latency_ms":579,"p99_latency_ms":621,"time_to_first_token_ms":143,"total_runtime_ms":527,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0289,"confidence_interval":0.9496,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2076,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":480,"p95_latency_ms":587,"p99_latency_ms":629,"time_to_first_token_ms":121,"total_runtime_ms":533,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0163,"confidence_interval":0.9436,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5381,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":446,"p95_latency_ms":545,"p99_latency_ms":585,"time_to_first_token_ms":145,"total_runtime_ms":496,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0238,"confidence_interval":0.9433,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9192,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":458,"p95_latency_ms":560,"p99_latency_ms":601,"time_to_first_token_ms":121,"total_runtime_ms":509,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0244,"confidence_interval":0.9479,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4139,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":441,"p95_latency_ms":539,"p99_latency_ms":579,"time_to_first_token_ms":149,"total_runtime_ms":490,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0177,"confidence_interval":0.9368,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.951,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":425,"p95_latency_ms":519,"p99_latency_ms":557,"time_to_first_token_ms":140,"total_runtime_ms":472,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0173,"confidence_interval":0.9457,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9024,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":306,"p95_latency_ms":374,"p99_latency_ms":401,"time_to_first_token_ms":80,"total_runtime_ms":340,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0223,"confidence_interval":0.9328,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7375,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":296,"p95_latency_ms":362,"p99_latency_ms":388,"time_to_first_token_ms":95,"total_runtime_ms":329,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0285,"confidence_interval":0.9389,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4344,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":296,"p95_latency_ms":361,"p99_latency_ms":388,"time_to_first_token_ms":99,"total_runtime_ms":329,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0227,"confidence_interval":0.9435,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4934,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":276,"p95_latency_ms":338,"p99_latency_ms":362,"time_to_first_token_ms":94,"total_runtime_ms":307,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0244,"confidence_interval":0.9354,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.905,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":253,"p95_latency_ms":309,"p99_latency_ms":331,"time_to_first_token_ms":99,"total_runtime_ms":281,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0131,"confidence_interval":0.9445,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6065,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":323,"p95_latency_ms":394,"p99_latency_ms":423,"time_to_first_token_ms":79,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.013,"confidence_interval":0.9484,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.625,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":345,"p95_latency_ms":421,"p99_latency_ms":452,"time_to_first_token_ms":85,"total_runtime_ms":383,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.01,"confidence_interval":0.9432,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9436,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":238,"p95_latency_ms":291,"p99_latency_ms":312,"time_to_first_token_ms":87,"total_runtime_ms":264,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.027,"confidence_interval":0.9367,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4381,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":277,"p95_latency_ms":339,"p99_latency_ms":364,"time_to_first_token_ms":88,"total_runtime_ms":308,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0268,"confidence_interval":0.9443,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7719,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":255,"p95_latency_ms":312,"p99_latency_ms":334,"time_to_first_token_ms":86,"total_runtime_ms":283,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0264,"confidence_interval":0.9461,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7227,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":748,"p95_latency_ms":914,"p99_latency_ms":981,"time_to_first_token_ms":177,"total_runtime_ms":831,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.019,"confidence_interval":0.9442,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6719,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":737,"p95_latency_ms":901,"p99_latency_ms":967,"time_to_first_token_ms":187,"total_runtime_ms":819,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.015,"confidence_interval":0.9327,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6626,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":781,"p95_latency_ms":955,"p99_latency_ms":1025,"time_to_first_token_ms":196,"total_runtime_ms":868,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0211,"confidence_interval":0.9369,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2516,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":808,"p95_latency_ms":987,"p99_latency_ms":1059,"time_to_first_token_ms":172,"total_runtime_ms":897,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0143,"confidence_interval":0.9301,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6063,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":792,"p95_latency_ms":968,"p99_latency_ms":1039,"time_to_first_token_ms":180,"total_runtime_ms":880,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0238,"confidence_interval":0.9424,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9773,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":771,"p95_latency_ms":942,"p99_latency_ms":1011,"time_to_first_token_ms":179,"total_runtime_ms":856,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.026,"confidence_interval":0.9499,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.014,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":753,"p95_latency_ms":920,"p99_latency_ms":987,"time_to_first_token_ms":185,"total_runtime_ms":837,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0293,"confidence_interval":0.9324,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3827,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":762,"p95_latency_ms":932,"p99_latency_ms":1000,"time_to_first_token_ms":168,"total_runtime_ms":847,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0241,"confidence_interval":0.9307,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":7.363,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":931,"time_to_first_token_ms":183,"total_runtime_ms":789,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0143,"confidence_interval":0.9353,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6384,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":739,"p95_latency_ms":903,"p99_latency_ms":969,"time_to_first_token_ms":185,"total_runtime_ms":821,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0265,"confidence_interval":0.9305,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9453,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":596,"time_to_first_token_ms":147,"total_runtime_ms":505,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.017,"confidence_interval":0.9371,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9582,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":499,"p95_latency_ms":610,"p99_latency_ms":654,"time_to_first_token_ms":127,"total_runtime_ms":554,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0146,"confidence_interval":0.9367,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7613,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":418,"p95_latency_ms":510,"p99_latency_ms":547,"time_to_first_token_ms":124,"total_runtime_ms":464,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0152,"confidence_interval":0.9495,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0072,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":413,"p95_latency_ms":505,"p99_latency_ms":542,"time_to_first_token_ms":146,"total_runtime_ms":459,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0189,"confidence_interval":0.9396,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3065,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":591,"time_to_first_token_ms":147,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0211,"confidence_interval":0.9409,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8047,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":465,"p95_latency_ms":568,"p99_latency_ms":610,"time_to_first_token_ms":137,"total_runtime_ms":517,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0215,"confidence_interval":0.9413,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2466,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":408,"p95_latency_ms":499,"p99_latency_ms":535,"time_to_first_token_ms":134,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0285,"confidence_interval":0.9498,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0631,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":500,"p95_latency_ms":611,"p99_latency_ms":656,"time_to_first_token_ms":136,"total_runtime_ms":556,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0229,"confidence_interval":0.9485,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4324,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":577,"time_to_first_token_ms":143,"total_runtime_ms":489,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.011,"confidence_interval":0.9328,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1229,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_strict_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":439,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":127,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0199,"confidence_interval":0.9379,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1061,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1100,"p95_latency_ms":1344,"p99_latency_ms":1442,"time_to_first_token_ms":213,"total_runtime_ms":1222,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0231,"confidence_interval":0.934,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5311,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1070,"p95_latency_ms":1307,"p99_latency_ms":1403,"time_to_first_token_ms":223,"total_runtime_ms":1189,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0275,"confidence_interval":0.9423,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0999,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1089,"p95_latency_ms":1331,"p99_latency_ms":1428,"time_to_first_token_ms":205,"total_runtime_ms":1210,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0244,"confidence_interval":0.9461,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1545,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1099,"p95_latency_ms":1344,"p99_latency_ms":1441,"time_to_first_token_ms":199,"total_runtime_ms":1221,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0245,"confidence_interval":0.9396,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1993,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1064,"p95_latency_ms":1300,"p99_latency_ms":1395,"time_to_first_token_ms":200,"total_runtime_ms":1182,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9436,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9764,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1067,"p95_latency_ms":1305,"p99_latency_ms":1400,"time_to_first_token_ms":192,"total_runtime_ms":1186,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0213,"confidence_interval":0.9459,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":5.678,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1057,"p95_latency_ms":1292,"p99_latency_ms":1385,"time_to_first_token_ms":201,"total_runtime_ms":1174,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0237,"confidence_interval":0.9391,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9971,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1125,"p95_latency_ms":1374,"p99_latency_ms":1474,"time_to_first_token_ms":190,"total_runtime_ms":1250,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0281,"confidence_interval":0.9395,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0554,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1130,"p95_latency_ms":1381,"p99_latency_ms":1482,"time_to_first_token_ms":202,"total_runtime_ms":1256,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0143,"confidence_interval":0.9374,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2595,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.912,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.088,"p50_latency_ms":1115,"p95_latency_ms":1363,"p99_latency_ms":1462,"time_to_first_token_ms":193,"total_runtime_ms":1239,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0235,"confidence_interval":0.9312,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3339,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":713,"p95_latency_ms":872,"p99_latency_ms":935,"time_to_first_token_ms":153,"total_runtime_ms":793,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0259,"confidence_interval":0.9392,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4267,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":746,"p95_latency_ms":912,"p99_latency_ms":979,"time_to_first_token_ms":161,"total_runtime_ms":829,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0269,"confidence_interval":0.9417,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7089,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":695,"p95_latency_ms":849,"p99_latency_ms":911,"time_to_first_token_ms":150,"total_runtime_ms":772,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0142,"confidence_interval":0.9496,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2326,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":681,"p95_latency_ms":832,"p99_latency_ms":893,"time_to_first_token_ms":164,"total_runtime_ms":757,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0296,"confidence_interval":0.9493,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6494,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":708,"p95_latency_ms":866,"p99_latency_ms":929,"time_to_first_token_ms":150,"total_runtime_ms":787,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0128,"confidence_interval":0.9442,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0688,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":747,"p95_latency_ms":913,"p99_latency_ms":979,"time_to_first_token_ms":143,"total_runtime_ms":830,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9497,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0595,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":747,"p95_latency_ms":913,"p99_latency_ms":980,"time_to_first_token_ms":170,"total_runtime_ms":830,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0229,"confidence_interval":0.9427,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9683,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":767,"p95_latency_ms":937,"p99_latency_ms":1005,"time_to_first_token_ms":164,"total_runtime_ms":852,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.028,"confidence_interval":0.9456,"validation_overhead_ms":94,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1994,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":742,"p95_latency_ms":907,"p99_latency_ms":973,"time_to_first_token_ms":167,"total_runtime_ms":825,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0296,"confidence_interval":0.932,"validation_overhead_ms":100,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3851,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.912,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.088,"p50_latency_ms":689,"p95_latency_ms":842,"p99_latency_ms":903,"time_to_first_token_ms":157,"total_runtime_ms":765,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0263,"confidence_interval":0.936,"validation_overhead_ms":88,"orchestration_overhead_ms":72,"event_dispatch_cost":6.275,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1810,"p95_latency_ms":2212,"p99_latency_ms":2373,"time_to_first_token_ms":285,"total_runtime_ms":2011,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0287,"confidence_interval":0.9345,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6468,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1817,"p95_latency_ms":2221,"p99_latency_ms":2382,"time_to_first_token_ms":285,"total_runtime_ms":2019,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.019,"confidence_interval":0.9359,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2471,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1826,"p95_latency_ms":2232,"p99_latency_ms":2395,"time_to_first_token_ms":287,"total_runtime_ms":2029,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0254,"confidence_interval":0.9301,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.539,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1811,"p95_latency_ms":2214,"p99_latency_ms":2375,"time_to_first_token_ms":289,"total_runtime_ms":2013,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9378,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1928,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1796,"p95_latency_ms":2195,"p99_latency_ms":2355,"time_to_first_token_ms":294,"total_runtime_ms":1996,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0195,"confidence_interval":0.9425,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.703,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1859,"p95_latency_ms":2273,"p99_latency_ms":2438,"time_to_first_token_ms":304,"total_runtime_ms":2066,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0147,"confidence_interval":0.937,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6525,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1812,"p95_latency_ms":2215,"p99_latency_ms":2376,"time_to_first_token_ms":293,"total_runtime_ms":2014,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0153,"confidence_interval":0.9307,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0052,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1888,"p95_latency_ms":2308,"p99_latency_ms":2476,"time_to_first_token_ms":297,"total_runtime_ms":2098,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0219,"confidence_interval":0.9387,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3647,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1853,"p95_latency_ms":2265,"p99_latency_ms":2430,"time_to_first_token_ms":302,"total_runtime_ms":2059,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0157,"confidence_interval":0.9398,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1778,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9008,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0992,"p50_latency_ms":1853,"p95_latency_ms":2264,"p99_latency_ms":2429,"time_to_first_token_ms":286,"total_runtime_ms":2058,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0271,"confidence_interval":0.948,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9345,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1212,"p95_latency_ms":1482,"p99_latency_ms":1590,"time_to_first_token_ms":248,"total_runtime_ms":1347,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0256,"confidence_interval":0.9387,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7283,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1177,"p95_latency_ms":1439,"p99_latency_ms":1544,"time_to_first_token_ms":247,"total_runtime_ms":1308,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0298,"confidence_interval":0.9437,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1239,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1224,"p95_latency_ms":1496,"p99_latency_ms":1605,"time_to_first_token_ms":250,"total_runtime_ms":1360,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0235,"confidence_interval":0.9439,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.288,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1239,"p95_latency_ms":1514,"p99_latency_ms":1624,"time_to_first_token_ms":247,"total_runtime_ms":1377,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.029,"confidence_interval":0.9339,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1107,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1225,"p95_latency_ms":1497,"p99_latency_ms":1606,"time_to_first_token_ms":247,"total_runtime_ms":1361,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0188,"confidence_interval":0.9308,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1066,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1212,"p95_latency_ms":1481,"p99_latency_ms":1589,"time_to_first_token_ms":257,"total_runtime_ms":1347,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.021,"confidence_interval":0.9422,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3547,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1242,"p95_latency_ms":1518,"p99_latency_ms":1629,"time_to_first_token_ms":239,"total_runtime_ms":1380,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0176,"confidence_interval":0.9355,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5842,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1230,"p95_latency_ms":1503,"p99_latency_ms":1612,"time_to_first_token_ms":245,"total_runtime_ms":1366,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0144,"confidence_interval":0.934,"validation_overhead_ms":107,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3828,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1185,"p95_latency_ms":1448,"p99_latency_ms":1554,"time_to_first_token_ms":264,"total_runtime_ms":1317,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0278,"confidence_interval":0.935,"validation_overhead_ms":113,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6801,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9008,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0992,"p50_latency_ms":1159,"p95_latency_ms":1416,"p99_latency_ms":1519,"time_to_first_token_ms":235,"total_runtime_ms":1288,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0118,"confidence_interval":0.9431,"validation_overhead_ms":101,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4075,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":847,"p95_latency_ms":1035,"p99_latency_ms":1110,"time_to_first_token_ms":184,"total_runtime_ms":941,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0209,"confidence_interval":0.938,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5487,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":850,"p95_latency_ms":1039,"p99_latency_ms":1115,"time_to_first_token_ms":166,"total_runtime_ms":945,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0206,"confidence_interval":0.9427,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6759,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":891,"p95_latency_ms":1089,"p99_latency_ms":1168,"time_to_first_token_ms":168,"total_runtime_ms":990,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.029,"confidence_interval":0.9316,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2373,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":157,"total_runtime_ms":889,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0168,"confidence_interval":0.9437,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8689,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":787,"p95_latency_ms":962,"p99_latency_ms":1032,"time_to_first_token_ms":176,"total_runtime_ms":875,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0265,"confidence_interval":0.9418,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9012,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":811,"p95_latency_ms":991,"p99_latency_ms":1063,"time_to_first_token_ms":159,"total_runtime_ms":901,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0141,"confidence_interval":0.9477,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9874,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":798,"p95_latency_ms":975,"p99_latency_ms":1046,"time_to_first_token_ms":179,"total_runtime_ms":886,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0295,"confidence_interval":0.9423,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1972,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":802,"p95_latency_ms":980,"p99_latency_ms":1052,"time_to_first_token_ms":170,"total_runtime_ms":891,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0197,"confidence_interval":0.9443,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4613,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":807,"p95_latency_ms":986,"p99_latency_ms":1058,"time_to_first_token_ms":158,"total_runtime_ms":896,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0162,"confidence_interval":0.946,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0013,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.92,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.08,"p50_latency_ms":824,"p95_latency_ms":1007,"p99_latency_ms":1081,"time_to_first_token_ms":169,"total_runtime_ms":916,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.029,"confidence_interval":0.943,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8334,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":533,"p95_latency_ms":651,"p99_latency_ms":699,"time_to_first_token_ms":124,"total_runtime_ms":592,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9444,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0252,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":542,"p95_latency_ms":663,"p99_latency_ms":711,"time_to_first_token_ms":119,"total_runtime_ms":602,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0167,"confidence_interval":0.9338,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0414,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":802,"time_to_first_token_ms":140,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0129,"confidence_interval":0.9425,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3034,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":613,"p95_latency_ms":750,"p99_latency_ms":804,"time_to_first_token_ms":124,"total_runtime_ms":681,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0229,"confidence_interval":0.9446,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1743,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":669,"time_to_first_token_ms":140,"total_runtime_ms":567,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0266,"confidence_interval":0.9476,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.266,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":509,"p95_latency_ms":623,"p99_latency_ms":668,"time_to_first_token_ms":119,"total_runtime_ms":566,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0223,"confidence_interval":0.9498,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9157,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":597,"p95_latency_ms":730,"p99_latency_ms":783,"time_to_first_token_ms":143,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0205,"confidence_interval":0.9348,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7348,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":129,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.026,"confidence_interval":0.9396,"validation_overhead_ms":94,"orchestration_overhead_ms":69,"event_dispatch_cost":6.337,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":772,"time_to_first_token_ms":119,"total_runtime_ms":654,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.017,"confidence_interval":0.9303,"validation_overhead_ms":100,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3254,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.92,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.08,"p50_latency_ms":533,"p95_latency_ms":652,"p99_latency_ms":699,"time_to_first_token_ms":114,"total_runtime_ms":592,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0176,"confidence_interval":0.9425,"validation_overhead_ms":88,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7845,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1442,"p95_latency_ms":1763,"p99_latency_ms":1891,"time_to_first_token_ms":255,"total_runtime_ms":1603,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.016,"confidence_interval":0.9362,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2049,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1391,"p95_latency_ms":1700,"p99_latency_ms":1824,"time_to_first_token_ms":246,"total_runtime_ms":1545,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0188,"confidence_interval":0.9439,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4144,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1436,"p95_latency_ms":1755,"p99_latency_ms":1883,"time_to_first_token_ms":227,"total_runtime_ms":1595,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.012,"confidence_interval":0.9462,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3083,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1415,"p95_latency_ms":1730,"p99_latency_ms":1856,"time_to_first_token_ms":250,"total_runtime_ms":1573,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0263,"confidence_interval":0.9313,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7204,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1364,"p95_latency_ms":1667,"p99_latency_ms":1788,"time_to_first_token_ms":258,"total_runtime_ms":1515,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0131,"confidence_interval":0.9323,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3514,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1400,"p95_latency_ms":1711,"p99_latency_ms":1835,"time_to_first_token_ms":251,"total_runtime_ms":1555,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0242,"confidence_interval":0.9365,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0386,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1375,"p95_latency_ms":1680,"p99_latency_ms":1802,"time_to_first_token_ms":255,"total_runtime_ms":1527,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0101,"confidence_interval":0.9398,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2125,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1363,"p95_latency_ms":1666,"p99_latency_ms":1787,"time_to_first_token_ms":245,"total_runtime_ms":1514,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.021,"confidence_interval":0.9365,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8733,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1370,"p95_latency_ms":1674,"p99_latency_ms":1796,"time_to_first_token_ms":231,"total_runtime_ms":1522,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0128,"confidence_interval":0.9372,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5885,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9088,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0912,"p50_latency_ms":1365,"p95_latency_ms":1669,"p99_latency_ms":1790,"time_to_first_token_ms":239,"total_runtime_ms":1517,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.014,"confidence_interval":0.9457,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.386,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":944,"p95_latency_ms":1154,"p99_latency_ms":1238,"time_to_first_token_ms":202,"total_runtime_ms":1049,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0276,"confidence_interval":0.9418,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":6.909,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":935,"p95_latency_ms":1143,"p99_latency_ms":1226,"time_to_first_token_ms":204,"total_runtime_ms":1039,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0175,"confidence_interval":0.9493,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0101,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":943,"p95_latency_ms":1153,"p99_latency_ms":1237,"time_to_first_token_ms":201,"total_runtime_ms":1048,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.016,"confidence_interval":0.9423,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2597,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":925,"p95_latency_ms":1131,"p99_latency_ms":1213,"time_to_first_token_ms":209,"total_runtime_ms":1028,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0216,"confidence_interval":0.9491,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4586,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":882,"p95_latency_ms":1078,"p99_latency_ms":1157,"time_to_first_token_ms":197,"total_runtime_ms":980,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0226,"confidence_interval":0.9441,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5292,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":939,"p95_latency_ms":1147,"p99_latency_ms":1231,"time_to_first_token_ms":186,"total_runtime_ms":1043,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0247,"confidence_interval":0.9327,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":6.871,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":948,"p95_latency_ms":1158,"p99_latency_ms":1243,"time_to_first_token_ms":205,"total_runtime_ms":1053,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0217,"confidence_interval":0.9419,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1125,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":943,"p95_latency_ms":1153,"p99_latency_ms":1237,"time_to_first_token_ms":187,"total_runtime_ms":1048,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.024,"confidence_interval":0.9333,"validation_overhead_ms":107,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5747,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":861,"p95_latency_ms":1052,"p99_latency_ms":1129,"time_to_first_token_ms":185,"total_runtime_ms":957,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0222,"confidence_interval":0.9472,"validation_overhead_ms":113,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3808,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9088,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0912,"p50_latency_ms":956,"p95_latency_ms":1168,"p99_latency_ms":1253,"time_to_first_token_ms":208,"total_runtime_ms":1062,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0241,"confidence_interval":0.9357,"validation_overhead_ms":101,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0261,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":654,"p95_latency_ms":799,"p99_latency_ms":857,"time_to_first_token_ms":148,"total_runtime_ms":726,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0287,"confidence_interval":0.9362,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9272,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":680,"p95_latency_ms":831,"p99_latency_ms":892,"time_to_first_token_ms":140,"total_runtime_ms":756,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0153,"confidence_interval":0.9305,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8797,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":647,"p95_latency_ms":791,"p99_latency_ms":848,"time_to_first_token_ms":164,"total_runtime_ms":719,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.024,"confidence_interval":0.9489,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1897,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":674,"p95_latency_ms":824,"p99_latency_ms":884,"time_to_first_token_ms":152,"total_runtime_ms":749,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0114,"confidence_interval":0.9417,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1572,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":903,"time_to_first_token_ms":164,"total_runtime_ms":765,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0136,"confidence_interval":0.9375,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9031,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":657,"p95_latency_ms":803,"p99_latency_ms":862,"time_to_first_token_ms":158,"total_runtime_ms":730,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0141,"confidence_interval":0.9315,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.691,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":652,"p95_latency_ms":796,"p99_latency_ms":854,"time_to_first_token_ms":162,"total_runtime_ms":724,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0176,"confidence_interval":0.9459,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0848,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":613,"p95_latency_ms":750,"p99_latency_ms":804,"time_to_first_token_ms":152,"total_runtime_ms":682,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0123,"confidence_interval":0.9364,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0041,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":624,"p95_latency_ms":763,"p99_latency_ms":818,"time_to_first_token_ms":156,"total_runtime_ms":693,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0271,"confidence_interval":0.9439,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7644,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.926,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.074,"p50_latency_ms":654,"p95_latency_ms":800,"p99_latency_ms":858,"time_to_first_token_ms":145,"total_runtime_ms":727,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0101,"confidence_interval":0.9401,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6211,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":117,"total_runtime_ms":534,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0178,"confidence_interval":0.9362,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8587,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":442,"p95_latency_ms":540,"p99_latency_ms":579,"time_to_first_token_ms":118,"total_runtime_ms":491,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0125,"confidence_interval":0.9398,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5199,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":412,"p95_latency_ms":503,"p99_latency_ms":540,"time_to_first_token_ms":110,"total_runtime_ms":458,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0277,"confidence_interval":0.9499,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2907,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":379,"p95_latency_ms":464,"p99_latency_ms":497,"time_to_first_token_ms":104,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0281,"confidence_interval":0.9325,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2043,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":409,"p95_latency_ms":500,"p99_latency_ms":536,"time_to_first_token_ms":102,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.9492,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0415,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":436,"p95_latency_ms":533,"p99_latency_ms":572,"time_to_first_token_ms":115,"total_runtime_ms":485,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0136,"confidence_interval":0.9387,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1021,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":483,"p95_latency_ms":590,"p99_latency_ms":633,"time_to_first_token_ms":110,"total_runtime_ms":536,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0209,"confidence_interval":0.9497,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1515,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":445,"p95_latency_ms":544,"p99_latency_ms":583,"time_to_first_token_ms":109,"total_runtime_ms":494,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0168,"confidence_interval":0.9484,"validation_overhead_ms":94,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0635,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":403,"p95_latency_ms":493,"p99_latency_ms":528,"time_to_first_token_ms":106,"total_runtime_ms":448,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0123,"confidence_interval":0.9476,"validation_overhead_ms":100,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3657,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.926,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.074,"p50_latency_ms":384,"p95_latency_ms":470,"p99_latency_ms":504,"time_to_first_token_ms":110,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.024,"confidence_interval":0.9394,"validation_overhead_ms":88,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6701,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1086,"p95_latency_ms":1327,"p99_latency_ms":1424,"time_to_first_token_ms":209,"total_runtime_ms":1206,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0242,"confidence_interval":0.9418,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0226,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1038,"p95_latency_ms":1269,"p99_latency_ms":1361,"time_to_first_token_ms":198,"total_runtime_ms":1153,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0217,"confidence_interval":0.9324,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5141,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1099,"p95_latency_ms":1343,"p99_latency_ms":1441,"time_to_first_token_ms":189,"total_runtime_ms":1221,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0159,"confidence_interval":0.9399,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5521,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1012,"p95_latency_ms":1237,"p99_latency_ms":1327,"time_to_first_token_ms":187,"total_runtime_ms":1125,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0283,"confidence_interval":0.9401,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":6.959,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1054,"p95_latency_ms":1289,"p99_latency_ms":1382,"time_to_first_token_ms":193,"total_runtime_ms":1171,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0275,"confidence_interval":0.9414,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5138,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1057,"p95_latency_ms":1292,"p99_latency_ms":1386,"time_to_first_token_ms":185,"total_runtime_ms":1175,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0133,"confidence_interval":0.9373,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6698,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1086,"p95_latency_ms":1327,"p99_latency_ms":1423,"time_to_first_token_ms":197,"total_runtime_ms":1206,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0191,"confidence_interval":0.9453,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.7194,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1080,"p95_latency_ms":1320,"p99_latency_ms":1416,"time_to_first_token_ms":205,"total_runtime_ms":1200,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0202,"confidence_interval":0.9455,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1425,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1087,"p95_latency_ms":1328,"p99_latency_ms":1425,"time_to_first_token_ms":192,"total_runtime_ms":1207,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.9348,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9405,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9148,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0852,"p50_latency_ms":1037,"p95_latency_ms":1268,"p99_latency_ms":1360,"time_to_first_token_ms":215,"total_runtime_ms":1152,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0254,"confidence_interval":0.9402,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5517,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":662,"p95_latency_ms":809,"p99_latency_ms":868,"time_to_first_token_ms":157,"total_runtime_ms":735,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0186,"confidence_interval":0.9359,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2011,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":690,"p95_latency_ms":843,"p99_latency_ms":904,"time_to_first_token_ms":162,"total_runtime_ms":766,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0112,"confidence_interval":0.9451,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8446,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":746,"p95_latency_ms":912,"p99_latency_ms":978,"time_to_first_token_ms":142,"total_runtime_ms":829,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0152,"confidence_interval":0.9409,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2632,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":699,"p95_latency_ms":855,"p99_latency_ms":917,"time_to_first_token_ms":163,"total_runtime_ms":777,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0263,"confidence_interval":0.9469,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6013,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":657,"p95_latency_ms":803,"p99_latency_ms":862,"time_to_first_token_ms":161,"total_runtime_ms":730,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0107,"confidence_interval":0.9379,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6837,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":867,"time_to_first_token_ms":159,"total_runtime_ms":734,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0231,"confidence_interval":0.9374,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3453,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":660,"p95_latency_ms":806,"p99_latency_ms":865,"time_to_first_token_ms":159,"total_runtime_ms":733,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0233,"confidence_interval":0.943,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":7.7188,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":728,"p95_latency_ms":890,"p99_latency_ms":954,"time_to_first_token_ms":154,"total_runtime_ms":809,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0151,"confidence_interval":0.9429,"validation_overhead_ms":107,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4911,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":747,"p95_latency_ms":913,"p99_latency_ms":980,"time_to_first_token_ms":172,"total_runtime_ms":830,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0124,"confidence_interval":0.9304,"validation_overhead_ms":113,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4075,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9148,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0852,"p50_latency_ms":675,"p95_latency_ms":825,"p99_latency_ms":885,"time_to_first_token_ms":168,"total_runtime_ms":750,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0208,"confidence_interval":0.9315,"validation_overhead_ms":101,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8008,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":552,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":125,"total_runtime_ms":613,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0222,"confidence_interval":0.9358,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7664,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":521,"p95_latency_ms":637,"p99_latency_ms":683,"time_to_first_token_ms":133,"total_runtime_ms":579,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0101,"confidence_interval":0.9319,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0456,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":144,"total_runtime_ms":611,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0115,"confidence_interval":0.9461,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5458,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":566,"p95_latency_ms":691,"p99_latency_ms":742,"time_to_first_token_ms":147,"total_runtime_ms":629,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0265,"confidence_interval":0.9455,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0539,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":565,"p95_latency_ms":690,"p99_latency_ms":740,"time_to_first_token_ms":146,"total_runtime_ms":627,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.012,"confidence_interval":0.9477,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5605,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":589,"p95_latency_ms":719,"p99_latency_ms":772,"time_to_first_token_ms":151,"total_runtime_ms":654,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0271,"confidence_interval":0.9365,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0121,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":756,"time_to_first_token_ms":153,"total_runtime_ms":641,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0181,"confidence_interval":0.9419,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9191,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":551,"p95_latency_ms":673,"p99_latency_ms":722,"time_to_first_token_ms":149,"total_runtime_ms":612,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0106,"confidence_interval":0.9333,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8042,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":542,"p95_latency_ms":662,"p99_latency_ms":710,"time_to_first_token_ms":128,"total_runtime_ms":602,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0121,"confidence_interval":0.9489,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1324,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.929,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.071,"p50_latency_ms":541,"p95_latency_ms":662,"p99_latency_ms":710,"time_to_first_token_ms":152,"total_runtime_ms":602,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0227,"confidence_interval":0.9492,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1845,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":335,"p95_latency_ms":409,"p99_latency_ms":439,"time_to_first_token_ms":102,"total_runtime_ms":372,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0194,"confidence_interval":0.9366,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5023,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":412,"p95_latency_ms":504,"p99_latency_ms":540,"time_to_first_token_ms":81,"total_runtime_ms":458,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0105,"confidence_interval":0.9474,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4714,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":370,"p95_latency_ms":453,"p99_latency_ms":486,"time_to_first_token_ms":97,"total_runtime_ms":411,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0287,"confidence_interval":0.9373,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3253,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":415,"p95_latency_ms":507,"p99_latency_ms":544,"time_to_first_token_ms":104,"total_runtime_ms":461,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0297,"confidence_interval":0.9314,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0283,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":537,"time_to_first_token_ms":92,"total_runtime_ms":455,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0128,"confidence_interval":0.9418,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7015,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":331,"p95_latency_ms":405,"p99_latency_ms":434,"time_to_first_token_ms":105,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.024,"confidence_interval":0.9448,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8578,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":366,"p95_latency_ms":447,"p99_latency_ms":480,"time_to_first_token_ms":107,"total_runtime_ms":407,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0132,"confidence_interval":0.9484,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":5.852,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":370,"p95_latency_ms":453,"p99_latency_ms":486,"time_to_first_token_ms":102,"total_runtime_ms":411,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.011,"confidence_interval":0.9353,"validation_overhead_ms":94,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6915,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":333,"p95_latency_ms":407,"p99_latency_ms":436,"time_to_first_token_ms":78,"total_runtime_ms":370,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0203,"confidence_interval":0.9375,"validation_overhead_ms":100,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1776,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.929,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.071,"p50_latency_ms":422,"p95_latency_ms":516,"p99_latency_ms":554,"time_to_first_token_ms":83,"total_runtime_ms":469,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0143,"confidence_interval":0.9473,"validation_overhead_ms":88,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5941,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":896,"p95_latency_ms":1095,"p99_latency_ms":1175,"time_to_first_token_ms":169,"total_runtime_ms":996,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.011,"confidence_interval":0.9409,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8055,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":914,"p95_latency_ms":1117,"p99_latency_ms":1198,"time_to_first_token_ms":168,"total_runtime_ms":1015,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0115,"confidence_interval":0.9384,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0702,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":941,"p95_latency_ms":1150,"p99_latency_ms":1234,"time_to_first_token_ms":180,"total_runtime_ms":1046,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0147,"confidence_interval":0.933,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7736,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":846,"p95_latency_ms":1034,"p99_latency_ms":1109,"time_to_first_token_ms":169,"total_runtime_ms":940,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0214,"confidence_interval":0.942,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2221,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":856,"p95_latency_ms":1047,"p99_latency_ms":1123,"time_to_first_token_ms":173,"total_runtime_ms":952,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0216,"confidence_interval":0.932,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4391,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":885,"p95_latency_ms":1081,"p99_latency_ms":1160,"time_to_first_token_ms":169,"total_runtime_ms":983,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0284,"confidence_interval":0.9484,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9106,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":900,"p95_latency_ms":1100,"p99_latency_ms":1180,"time_to_first_token_ms":183,"total_runtime_ms":1000,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0112,"confidence_interval":0.9371,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6047,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":837,"p95_latency_ms":1023,"p99_latency_ms":1098,"time_to_first_token_ms":163,"total_runtime_ms":930,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0135,"confidence_interval":0.9426,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5481,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":877,"p95_latency_ms":1072,"p99_latency_ms":1150,"time_to_first_token_ms":189,"total_runtime_ms":974,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0227,"confidence_interval":0.9331,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2318,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.9178,"parse_success_rate":0.922,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.0822,"p50_latency_ms":878,"p95_latency_ms":1073,"p99_latency_ms":1151,"time_to_first_token_ms":172,"total_runtime_ms":975,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0121,"confidence_interval":0.9329,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9631,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":634,"p95_latency_ms":775,"p99_latency_ms":831,"time_to_first_token_ms":143,"total_runtime_ms":704,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0299,"confidence_interval":0.9341,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9606,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":540,"p95_latency_ms":660,"p99_latency_ms":708,"time_to_first_token_ms":130,"total_runtime_ms":600,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0137,"confidence_interval":0.9455,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8108,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":118,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0207,"confidence_interval":0.9425,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4788,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":552,"p95_latency_ms":675,"p99_latency_ms":724,"time_to_first_token_ms":120,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0148,"confidence_interval":0.935,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8238,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":626,"p95_latency_ms":765,"p99_latency_ms":820,"time_to_first_token_ms":117,"total_runtime_ms":695,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0162,"confidence_interval":0.9349,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8838,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":635,"p95_latency_ms":776,"p99_latency_ms":833,"time_to_first_token_ms":152,"total_runtime_ms":706,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.011,"confidence_interval":0.9388,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.68,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":577,"p95_latency_ms":706,"p99_latency_ms":757,"time_to_first_token_ms":140,"total_runtime_ms":642,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0162,"confidence_interval":0.9316,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9258,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":123,"total_runtime_ms":650,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0253,"confidence_interval":0.943,"validation_overhead_ms":107,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0829,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":625,"p95_latency_ms":763,"p99_latency_ms":819,"time_to_first_token_ms":131,"total_runtime_ms":694,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0277,"confidence_interval":0.9373,"validation_overhead_ms":113,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1146,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"extract_nested_json","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.9178,"parse_success_rate":0.952,"retry_rate":0.09,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.0822,"p50_latency_ms":619,"p95_latency_ms":756,"p99_latency_ms":811,"time_to_first_token_ms":132,"total_runtime_ms":687,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0227,"confidence_interval":0.9424,"validation_overhead_ms":101,"orchestration_overhead_ms":65,"event_dispatch_cost":7.619,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":815,"p95_latency_ms":996,"p99_latency_ms":1069,"time_to_first_token_ms":208,"total_runtime_ms":906,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0274,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8318,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":832,"p95_latency_ms":1017,"p99_latency_ms":1091,"time_to_first_token_ms":213,"total_runtime_ms":925,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5128,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":750,"p95_latency_ms":917,"p99_latency_ms":984,"time_to_first_token_ms":209,"total_runtime_ms":834,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0103,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9809,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":816,"p95_latency_ms":998,"p99_latency_ms":1070,"time_to_first_token_ms":215,"total_runtime_ms":907,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0247,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.358,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":831,"p95_latency_ms":1016,"p99_latency_ms":1090,"time_to_first_token_ms":194,"total_runtime_ms":923,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0289,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8607,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":811,"p95_latency_ms":991,"p99_latency_ms":1063,"time_to_first_token_ms":211,"total_runtime_ms":901,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0242,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3208,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":804,"p95_latency_ms":983,"p99_latency_ms":1054,"time_to_first_token_ms":193,"total_runtime_ms":893,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0137,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3862,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1014,"p99_latency_ms":1088,"time_to_first_token_ms":219,"total_runtime_ms":922,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0196,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4945,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":831,"p95_latency_ms":1015,"p99_latency_ms":1089,"time_to_first_token_ms":193,"total_runtime_ms":923,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.477,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":829,"p95_latency_ms":1013,"p99_latency_ms":1087,"time_to_first_token_ms":214,"total_runtime_ms":921,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0118,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8844,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":152,"total_runtime_ms":436,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2255,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":144,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0228,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0082,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":375,"p95_latency_ms":458,"p99_latency_ms":492,"time_to_first_token_ms":167,"total_runtime_ms":417,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0185,"confidence_interval":0.9336,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.432,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":468,"p95_latency_ms":572,"p99_latency_ms":614,"time_to_first_token_ms":172,"total_runtime_ms":520,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0261,"confidence_interval":0.9369,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6994,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":420,"p95_latency_ms":514,"p99_latency_ms":551,"time_to_first_token_ms":143,"total_runtime_ms":467,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0206,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2461,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":149,"total_runtime_ms":476,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0128,"confidence_interval":0.944,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6792,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":429,"p95_latency_ms":524,"p99_latency_ms":562,"time_to_first_token_ms":160,"total_runtime_ms":476,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0151,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1889,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":470,"p95_latency_ms":574,"p99_latency_ms":616,"time_to_first_token_ms":146,"total_runtime_ms":522,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0265,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9352,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":163,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0135,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6708,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":591,"time_to_first_token_ms":149,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.027,"confidence_interval":0.9423,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3803,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1357,"p95_latency_ms":1659,"p99_latency_ms":1779,"time_to_first_token_ms":283,"total_runtime_ms":1508,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0221,"confidence_interval":0.9304,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5654,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1317,"p95_latency_ms":1609,"p99_latency_ms":1726,"time_to_first_token_ms":304,"total_runtime_ms":1463,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0166,"confidence_interval":0.939,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9273,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1328,"p95_latency_ms":1624,"p99_latency_ms":1742,"time_to_first_token_ms":309,"total_runtime_ms":1476,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0143,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5875,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1333,"p95_latency_ms":1629,"p99_latency_ms":1747,"time_to_first_token_ms":278,"total_runtime_ms":1481,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.027,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4536,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1305,"p95_latency_ms":1595,"p99_latency_ms":1711,"time_to_first_token_ms":297,"total_runtime_ms":1450,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0104,"confidence_interval":0.9362,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1256,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1307,"p95_latency_ms":1597,"p99_latency_ms":1713,"time_to_first_token_ms":283,"total_runtime_ms":1452,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0122,"confidence_interval":0.9472,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4025,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1287,"p95_latency_ms":1573,"p99_latency_ms":1688,"time_to_first_token_ms":280,"total_runtime_ms":1430,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0149,"confidence_interval":0.9395,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2691,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1340,"p95_latency_ms":1638,"p99_latency_ms":1757,"time_to_first_token_ms":301,"total_runtime_ms":1489,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0133,"confidence_interval":0.9495,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2365,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1369,"p95_latency_ms":1674,"p99_latency_ms":1795,"time_to_first_token_ms":305,"total_runtime_ms":1521,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0124,"confidence_interval":0.9331,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5604,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1316,"p95_latency_ms":1608,"p99_latency_ms":1725,"time_to_first_token_ms":312,"total_runtime_ms":1462,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0116,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8529,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":663,"p95_latency_ms":810,"p99_latency_ms":869,"time_to_first_token_ms":251,"total_runtime_ms":736,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0289,"confidence_interval":0.9362,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3405,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":697,"p95_latency_ms":852,"p99_latency_ms":914,"time_to_first_token_ms":238,"total_runtime_ms":774,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0259,"confidence_interval":0.9314,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8309,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":735,"p95_latency_ms":898,"p99_latency_ms":963,"time_to_first_token_ms":265,"total_runtime_ms":816,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0268,"confidence_interval":0.9429,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2029,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":725,"p95_latency_ms":886,"p99_latency_ms":951,"time_to_first_token_ms":244,"total_runtime_ms":806,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0224,"confidence_interval":0.9474,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6928,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":681,"p95_latency_ms":832,"p99_latency_ms":893,"time_to_first_token_ms":237,"total_runtime_ms":757,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0271,"confidence_interval":0.9363,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3272,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":718,"p95_latency_ms":877,"p99_latency_ms":941,"time_to_first_token_ms":246,"total_runtime_ms":798,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0275,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.7134,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":713,"p95_latency_ms":871,"p99_latency_ms":934,"time_to_first_token_ms":252,"total_runtime_ms":792,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.029,"confidence_interval":0.9376,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.777,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":683,"p95_latency_ms":834,"p99_latency_ms":895,"time_to_first_token_ms":257,"total_runtime_ms":759,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0241,"confidence_interval":0.946,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0499,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":726,"p95_latency_ms":887,"p99_latency_ms":952,"time_to_first_token_ms":233,"total_runtime_ms":806,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0144,"confidence_interval":0.9318,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0819,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":664,"p95_latency_ms":812,"p99_latency_ms":871,"time_to_first_token_ms":260,"total_runtime_ms":738,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0271,"confidence_interval":0.9478,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6868,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":624,"p95_latency_ms":763,"p99_latency_ms":818,"time_to_first_token_ms":161,"total_runtime_ms":693,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0282,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3371,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":641,"p95_latency_ms":783,"p99_latency_ms":840,"time_to_first_token_ms":189,"total_runtime_ms":712,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0245,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4179,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":851,"time_to_first_token_ms":167,"total_runtime_ms":721,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0228,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1516,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":571,"p95_latency_ms":698,"p99_latency_ms":749,"time_to_first_token_ms":164,"total_runtime_ms":634,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0227,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8942,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":599,"p95_latency_ms":733,"p99_latency_ms":786,"time_to_first_token_ms":168,"total_runtime_ms":666,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0279,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6481,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":579,"p95_latency_ms":708,"p99_latency_ms":759,"time_to_first_token_ms":189,"total_runtime_ms":644,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0268,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1071,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":756,"time_to_first_token_ms":176,"total_runtime_ms":640,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0114,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7247,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":575,"p95_latency_ms":703,"p99_latency_ms":754,"time_to_first_token_ms":173,"total_runtime_ms":639,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0184,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7651,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":654,"p95_latency_ms":799,"p99_latency_ms":857,"time_to_first_token_ms":183,"total_runtime_ms":726,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0278,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.45,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":645,"p95_latency_ms":789,"p99_latency_ms":846,"time_to_first_token_ms":191,"total_runtime_ms":717,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.027,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3451,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":362,"p95_latency_ms":442,"p99_latency_ms":474,"time_to_first_token_ms":143,"total_runtime_ms":402,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0155,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3251,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":461,"p99_latency_ms":495,"time_to_first_token_ms":114,"total_runtime_ms":419,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.015,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3764,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":382,"p95_latency_ms":467,"p99_latency_ms":500,"time_to_first_token_ms":144,"total_runtime_ms":424,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0177,"confidence_interval":0.9303,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3268,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":283,"p95_latency_ms":346,"p99_latency_ms":371,"time_to_first_token_ms":141,"total_runtime_ms":314,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0155,"confidence_interval":0.9419,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8688,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":353,"p95_latency_ms":432,"p99_latency_ms":463,"time_to_first_token_ms":136,"total_runtime_ms":393,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0184,"confidence_interval":0.9338,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5112,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":372,"p95_latency_ms":455,"p99_latency_ms":488,"time_to_first_token_ms":111,"total_runtime_ms":414,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0238,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.799,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":366,"p95_latency_ms":447,"p99_latency_ms":479,"time_to_first_token_ms":115,"total_runtime_ms":406,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0233,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8948,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":295,"p95_latency_ms":360,"p99_latency_ms":386,"time_to_first_token_ms":133,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0222,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4548,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":313,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":121,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0256,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9595,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":354,"p95_latency_ms":432,"p99_latency_ms":464,"time_to_first_token_ms":117,"total_runtime_ms":393,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0293,"confidence_interval":0.9455,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5979,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1047,"p95_latency_ms":1280,"p99_latency_ms":1373,"time_to_first_token_ms":239,"total_runtime_ms":1163,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.014,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3727,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1047,"p95_latency_ms":1280,"p99_latency_ms":1373,"time_to_first_token_ms":242,"total_runtime_ms":1164,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0296,"confidence_interval":0.9384,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4052,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":988,"p95_latency_ms":1208,"p99_latency_ms":1295,"time_to_first_token_ms":228,"total_runtime_ms":1098,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0111,"confidence_interval":0.9461,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9462,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1008,"p95_latency_ms":1232,"p99_latency_ms":1322,"time_to_first_token_ms":227,"total_runtime_ms":1120,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0193,"confidence_interval":0.9439,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9659,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1016,"p95_latency_ms":1241,"p99_latency_ms":1332,"time_to_first_token_ms":242,"total_runtime_ms":1128,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0149,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2095,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1062,"p95_latency_ms":1298,"p99_latency_ms":1392,"time_to_first_token_ms":235,"total_runtime_ms":1180,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0262,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5684,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1019,"p95_latency_ms":1245,"p99_latency_ms":1336,"time_to_first_token_ms":239,"total_runtime_ms":1132,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0122,"confidence_interval":0.9335,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2045,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":977,"p95_latency_ms":1194,"p99_latency_ms":1281,"time_to_first_token_ms":229,"total_runtime_ms":1086,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0226,"confidence_interval":0.9308,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5238,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":971,"p95_latency_ms":1186,"p99_latency_ms":1272,"time_to_first_token_ms":246,"total_runtime_ms":1078,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0188,"confidence_interval":0.9411,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1548,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1051,"p95_latency_ms":1284,"p99_latency_ms":1377,"time_to_first_token_ms":254,"total_runtime_ms":1167,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0218,"confidence_interval":0.9347,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3209,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":504,"p95_latency_ms":616,"p99_latency_ms":661,"time_to_first_token_ms":195,"total_runtime_ms":560,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0261,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0145,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":505,"p95_latency_ms":617,"p99_latency_ms":662,"time_to_first_token_ms":182,"total_runtime_ms":561,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0111,"confidence_interval":0.9466,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0639,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":520,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":202,"total_runtime_ms":578,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0183,"confidence_interval":0.9425,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6821,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":563,"p95_latency_ms":688,"p99_latency_ms":738,"time_to_first_token_ms":200,"total_runtime_ms":625,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.9335,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2034,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":181,"total_runtime_ms":636,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0198,"confidence_interval":0.9397,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4496,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":554,"p95_latency_ms":677,"p99_latency_ms":726,"time_to_first_token_ms":191,"total_runtime_ms":616,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0127,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1424,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":498,"p95_latency_ms":609,"p99_latency_ms":653,"time_to_first_token_ms":195,"total_runtime_ms":553,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.019,"confidence_interval":0.9346,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6348,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":543,"p95_latency_ms":663,"p99_latency_ms":712,"time_to_first_token_ms":193,"total_runtime_ms":603,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0147,"confidence_interval":0.9407,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7905,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":540,"p95_latency_ms":660,"p99_latency_ms":708,"time_to_first_token_ms":194,"total_runtime_ms":600,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.016,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2471,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":529,"p95_latency_ms":647,"p99_latency_ms":694,"time_to_first_token_ms":198,"total_runtime_ms":588,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0107,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8114,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":495,"p95_latency_ms":605,"p99_latency_ms":649,"time_to_first_token_ms":161,"total_runtime_ms":550,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0228,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.56,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":503,"p95_latency_ms":615,"p99_latency_ms":659,"time_to_first_token_ms":152,"total_runtime_ms":559,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0202,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7373,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":528,"p95_latency_ms":645,"p99_latency_ms":692,"time_to_first_token_ms":143,"total_runtime_ms":587,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0105,"confidence_interval":0.9351,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0686,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":474,"p95_latency_ms":579,"p99_latency_ms":621,"time_to_first_token_ms":141,"total_runtime_ms":526,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0238,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2892,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":158,"total_runtime_ms":575,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6143,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":519,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":136,"total_runtime_ms":577,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0203,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6984,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":164,"total_runtime_ms":536,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0173,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.287,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":138,"total_runtime_ms":535,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0203,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.83,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":515,"p95_latency_ms":629,"p99_latency_ms":675,"time_to_first_token_ms":145,"total_runtime_ms":572,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0215,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2662,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":448,"p95_latency_ms":548,"p99_latency_ms":588,"time_to_first_token_ms":157,"total_runtime_ms":498,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.9465,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5615,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":275,"p95_latency_ms":336,"p99_latency_ms":360,"time_to_first_token_ms":97,"total_runtime_ms":305,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0245,"confidence_interval":0.9319,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1169,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":266,"p95_latency_ms":326,"p99_latency_ms":349,"time_to_first_token_ms":109,"total_runtime_ms":296,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0145,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0979,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":261,"p95_latency_ms":319,"p99_latency_ms":342,"time_to_first_token_ms":100,"total_runtime_ms":290,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0277,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0983,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":223,"p95_latency_ms":273,"p99_latency_ms":293,"time_to_first_token_ms":120,"total_runtime_ms":248,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0241,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8713,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":288,"p95_latency_ms":352,"p99_latency_ms":378,"time_to_first_token_ms":103,"total_runtime_ms":320,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0155,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.133,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":301,"p95_latency_ms":368,"p99_latency_ms":395,"time_to_first_token_ms":120,"total_runtime_ms":335,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0279,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4574,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":256,"p95_latency_ms":313,"p99_latency_ms":336,"time_to_first_token_ms":90,"total_runtime_ms":285,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0183,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4622,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":236,"p95_latency_ms":289,"p99_latency_ms":310,"time_to_first_token_ms":96,"total_runtime_ms":263,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0167,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6895,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":279,"p95_latency_ms":341,"p99_latency_ms":365,"time_to_first_token_ms":117,"total_runtime_ms":310,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0213,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9643,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":302,"p95_latency_ms":369,"p99_latency_ms":395,"time_to_first_token_ms":89,"total_runtime_ms":335,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.01,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9213,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":948,"time_to_first_token_ms":187,"total_runtime_ms":804,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0246,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7885,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":806,"p95_latency_ms":985,"p99_latency_ms":1056,"time_to_first_token_ms":184,"total_runtime_ms":895,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0173,"confidence_interval":0.9344,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2103,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":717,"p95_latency_ms":876,"p99_latency_ms":940,"time_to_first_token_ms":196,"total_runtime_ms":797,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0282,"confidence_interval":0.9381,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5894,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":805,"p95_latency_ms":984,"p99_latency_ms":1056,"time_to_first_token_ms":214,"total_runtime_ms":895,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0291,"confidence_interval":0.9344,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1585,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":726,"p95_latency_ms":888,"p99_latency_ms":952,"time_to_first_token_ms":183,"total_runtime_ms":807,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0279,"confidence_interval":0.9485,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0237,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1049,"time_to_first_token_ms":190,"total_runtime_ms":889,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0273,"confidence_interval":0.9322,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0546,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":769,"p95_latency_ms":940,"p99_latency_ms":1008,"time_to_first_token_ms":214,"total_runtime_ms":854,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0187,"confidence_interval":0.9384,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2589,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":754,"p95_latency_ms":921,"p99_latency_ms":988,"time_to_first_token_ms":192,"total_runtime_ms":838,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0208,"confidence_interval":0.9316,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6639,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":769,"p95_latency_ms":940,"p99_latency_ms":1008,"time_to_first_token_ms":217,"total_runtime_ms":855,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0216,"confidence_interval":0.9384,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5989,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":732,"p95_latency_ms":894,"p99_latency_ms":959,"time_to_first_token_ms":193,"total_runtime_ms":813,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.028,"confidence_interval":0.9373,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7293,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":461,"p95_latency_ms":563,"p99_latency_ms":604,"time_to_first_token_ms":148,"total_runtime_ms":512,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0281,"confidence_interval":0.9436,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4411,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":433,"p95_latency_ms":529,"p99_latency_ms":567,"time_to_first_token_ms":144,"total_runtime_ms":481,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.015,"confidence_interval":0.939,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.86,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":424,"p95_latency_ms":519,"p99_latency_ms":556,"time_to_first_token_ms":159,"total_runtime_ms":472,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0223,"confidence_interval":0.9471,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6154,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":398,"p95_latency_ms":486,"p99_latency_ms":522,"time_to_first_token_ms":154,"total_runtime_ms":442,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0208,"confidence_interval":0.9398,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0388,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":551,"p99_latency_ms":591,"time_to_first_token_ms":141,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.021,"confidence_interval":0.9315,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2626,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":405,"p95_latency_ms":495,"p99_latency_ms":531,"time_to_first_token_ms":147,"total_runtime_ms":450,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0114,"confidence_interval":0.9485,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1047,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":366,"p95_latency_ms":447,"p99_latency_ms":480,"time_to_first_token_ms":167,"total_runtime_ms":407,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0265,"confidence_interval":0.9391,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1988,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":385,"p95_latency_ms":470,"p99_latency_ms":504,"time_to_first_token_ms":153,"total_runtime_ms":427,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.018,"confidence_interval":0.9473,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5441,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":371,"p95_latency_ms":454,"p99_latency_ms":487,"time_to_first_token_ms":168,"total_runtime_ms":413,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0291,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8788,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":448,"p95_latency_ms":548,"p99_latency_ms":587,"time_to_first_token_ms":169,"total_runtime_ms":498,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0239,"confidence_interval":0.9371,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1158,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":458,"p95_latency_ms":559,"p99_latency_ms":600,"time_to_first_token_ms":136,"total_runtime_ms":508,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0266,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0523,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":411,"p95_latency_ms":502,"p99_latency_ms":538,"time_to_first_token_ms":138,"total_runtime_ms":456,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0197,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9387,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":419,"p95_latency_ms":513,"p99_latency_ms":550,"time_to_first_token_ms":154,"total_runtime_ms":466,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0209,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6971,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":123,"total_runtime_ms":456,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0143,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8358,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":368,"p95_latency_ms":449,"p99_latency_ms":482,"time_to_first_token_ms":127,"total_runtime_ms":409,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0216,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1407,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":442,"p95_latency_ms":541,"p99_latency_ms":580,"time_to_first_token_ms":125,"total_runtime_ms":492,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0147,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3398,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":144,"total_runtime_ms":476,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0271,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3954,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":427,"p95_latency_ms":522,"p99_latency_ms":560,"time_to_first_token_ms":138,"total_runtime_ms":474,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0283,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3362,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":406,"p95_latency_ms":496,"p99_latency_ms":533,"time_to_first_token_ms":132,"total_runtime_ms":451,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0288,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7001,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":548,"time_to_first_token_ms":150,"total_runtime_ms":464,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0259,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.474,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":216,"p95_latency_ms":265,"p99_latency_ms":284,"time_to_first_token_ms":78,"total_runtime_ms":241,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0161,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0947,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":203,"p95_latency_ms":249,"p99_latency_ms":267,"time_to_first_token_ms":80,"total_runtime_ms":226,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0113,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.164,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":229,"p95_latency_ms":280,"p99_latency_ms":300,"time_to_first_token_ms":92,"total_runtime_ms":254,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0193,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.934,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":272,"p95_latency_ms":332,"p99_latency_ms":356,"time_to_first_token_ms":88,"total_runtime_ms":302,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0148,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0186,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":188,"p95_latency_ms":230,"p99_latency_ms":247,"time_to_first_token_ms":101,"total_runtime_ms":209,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0133,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8387,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":194,"p95_latency_ms":237,"p99_latency_ms":254,"time_to_first_token_ms":85,"total_runtime_ms":215,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0299,"confidence_interval":0.9326,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2253,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":241,"p95_latency_ms":295,"p99_latency_ms":316,"time_to_first_token_ms":100,"total_runtime_ms":268,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0138,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4173,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":204,"p95_latency_ms":249,"p99_latency_ms":267,"time_to_first_token_ms":107,"total_runtime_ms":227,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0188,"confidence_interval":0.9315,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.079,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":275,"p95_latency_ms":336,"p99_latency_ms":361,"time_to_first_token_ms":90,"total_runtime_ms":306,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0285,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3841,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":222,"p95_latency_ms":271,"p99_latency_ms":291,"time_to_first_token_ms":75,"total_runtime_ms":246,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0217,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5794,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":657,"p95_latency_ms":802,"p99_latency_ms":861,"time_to_first_token_ms":179,"total_runtime_ms":730,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.023,"confidence_interval":0.9446,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1983,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":691,"p95_latency_ms":845,"p99_latency_ms":907,"time_to_first_token_ms":178,"total_runtime_ms":768,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0209,"confidence_interval":0.9489,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.685,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":694,"p95_latency_ms":849,"p99_latency_ms":910,"time_to_first_token_ms":178,"total_runtime_ms":772,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0225,"confidence_interval":0.9356,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3792,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":700,"p95_latency_ms":855,"p99_latency_ms":918,"time_to_first_token_ms":181,"total_runtime_ms":778,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0267,"confidence_interval":0.9405,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.003,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":678,"p95_latency_ms":829,"p99_latency_ms":889,"time_to_first_token_ms":167,"total_runtime_ms":753,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0246,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6147,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":851,"time_to_first_token_ms":189,"total_runtime_ms":721,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.016,"confidence_interval":0.9485,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0028,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":668,"p95_latency_ms":817,"p99_latency_ms":876,"time_to_first_token_ms":191,"total_runtime_ms":743,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0263,"confidence_interval":0.936,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.451,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":650,"p95_latency_ms":794,"p99_latency_ms":852,"time_to_first_token_ms":178,"total_runtime_ms":722,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0255,"confidence_interval":0.9375,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4222,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":638,"p95_latency_ms":780,"p99_latency_ms":836,"time_to_first_token_ms":172,"total_runtime_ms":709,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.026,"confidence_interval":0.9399,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1064,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":618,"p95_latency_ms":756,"p99_latency_ms":811,"time_to_first_token_ms":170,"total_runtime_ms":687,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0109,"confidence_interval":0.9399,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1217,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":377,"p99_latency_ms":404,"time_to_first_token_ms":128,"total_runtime_ms":343,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0176,"confidence_interval":0.9328,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.001,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":339,"p95_latency_ms":414,"p99_latency_ms":445,"time_to_first_token_ms":131,"total_runtime_ms":377,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0169,"confidence_interval":0.9408,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.541,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":314,"p95_latency_ms":384,"p99_latency_ms":412,"time_to_first_token_ms":149,"total_runtime_ms":349,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0234,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0086,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":357,"p95_latency_ms":437,"p99_latency_ms":469,"time_to_first_token_ms":138,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0155,"confidence_interval":0.9443,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0018,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":317,"p95_latency_ms":388,"p99_latency_ms":416,"time_to_first_token_ms":149,"total_runtime_ms":352,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.019,"confidence_interval":0.9305,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.944,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":477,"p99_latency_ms":512,"time_to_first_token_ms":118,"total_runtime_ms":434,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0248,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4448,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":398,"p95_latency_ms":486,"p99_latency_ms":522,"time_to_first_token_ms":133,"total_runtime_ms":442,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0197,"confidence_interval":0.9321,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4008,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":313,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":119,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0234,"confidence_interval":0.937,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4818,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":302,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":140,"total_runtime_ms":336,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0155,"confidence_interval":0.9338,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4197,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"summarize_contextual","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":350,"p95_latency_ms":428,"p99_latency_ms":460,"time_to_first_token_ms":125,"total_runtime_ms":389,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0108,"confidence_interval":0.9419,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2151,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1044,"p95_latency_ms":1276,"p99_latency_ms":1369,"time_to_first_token_ms":206,"total_runtime_ms":1160,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0151,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0362,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1043,"p95_latency_ms":1275,"p99_latency_ms":1367,"time_to_first_token_ms":218,"total_runtime_ms":1159,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0195,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2703,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1000,"p95_latency_ms":1222,"p99_latency_ms":1311,"time_to_first_token_ms":210,"total_runtime_ms":1111,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0125,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3919,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1019,"p95_latency_ms":1245,"p99_latency_ms":1336,"time_to_first_token_ms":200,"total_runtime_ms":1132,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.021,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.436,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":951,"p95_latency_ms":1162,"p99_latency_ms":1247,"time_to_first_token_ms":196,"total_runtime_ms":1056,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9534,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":974,"p95_latency_ms":1190,"p99_latency_ms":1276,"time_to_first_token_ms":197,"total_runtime_ms":1082,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0253,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2661,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":966,"p95_latency_ms":1180,"p99_latency_ms":1266,"time_to_first_token_ms":205,"total_runtime_ms":1073,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0278,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2679,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1002,"p95_latency_ms":1225,"p99_latency_ms":1314,"time_to_first_token_ms":222,"total_runtime_ms":1114,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0268,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6665,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":937,"p95_latency_ms":1145,"p99_latency_ms":1228,"time_to_first_token_ms":206,"total_runtime_ms":1041,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0183,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1104,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1027,"p95_latency_ms":1255,"p99_latency_ms":1346,"time_to_first_token_ms":209,"total_runtime_ms":1141,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0146,"confidence_interval":0.9475,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3002,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":638,"p95_latency_ms":780,"p99_latency_ms":837,"time_to_first_token_ms":162,"total_runtime_ms":709,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9369,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0371,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":152,"total_runtime_ms":741,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0258,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2813,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":664,"p95_latency_ms":811,"p99_latency_ms":870,"time_to_first_token_ms":163,"total_runtime_ms":737,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0273,"confidence_interval":0.9468,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6156,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":850,"time_to_first_token_ms":145,"total_runtime_ms":721,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0175,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1792,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":802,"time_to_first_token_ms":144,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0183,"confidence_interval":0.933,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.039,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":604,"p95_latency_ms":739,"p99_latency_ms":792,"time_to_first_token_ms":161,"total_runtime_ms":672,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0206,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1583,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":636,"p95_latency_ms":778,"p99_latency_ms":834,"time_to_first_token_ms":150,"total_runtime_ms":707,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0183,"confidence_interval":0.9321,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4318,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":617,"p95_latency_ms":754,"p99_latency_ms":809,"time_to_first_token_ms":152,"total_runtime_ms":686,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0148,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9629,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":626,"p95_latency_ms":766,"p99_latency_ms":821,"time_to_first_token_ms":175,"total_runtime_ms":696,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.027,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.472,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":627,"p95_latency_ms":766,"p99_latency_ms":822,"time_to_first_token_ms":162,"total_runtime_ms":696,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0175,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2898,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1673,"p95_latency_ms":2044,"p99_latency_ms":2193,"time_to_first_token_ms":281,"total_runtime_ms":1858,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.946,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0115,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1634,"p95_latency_ms":1997,"p99_latency_ms":2142,"time_to_first_token_ms":298,"total_runtime_ms":1816,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0176,"confidence_interval":0.9497,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2408,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1664,"p95_latency_ms":2034,"p99_latency_ms":2182,"time_to_first_token_ms":293,"total_runtime_ms":1849,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0284,"confidence_interval":0.9473,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.593,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1675,"p95_latency_ms":2047,"p99_latency_ms":2196,"time_to_first_token_ms":280,"total_runtime_ms":1861,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0107,"confidence_interval":0.934,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4405,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1646,"p95_latency_ms":2011,"p99_latency_ms":2158,"time_to_first_token_ms":296,"total_runtime_ms":1829,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0111,"confidence_interval":0.9476,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7794,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1685,"p95_latency_ms":2059,"p99_latency_ms":2209,"time_to_first_token_ms":285,"total_runtime_ms":1872,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9495,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7306,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1620,"p95_latency_ms":1980,"p99_latency_ms":2124,"time_to_first_token_ms":298,"total_runtime_ms":1800,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0226,"confidence_interval":0.9389,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.176,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1622,"p95_latency_ms":1982,"p99_latency_ms":2126,"time_to_first_token_ms":281,"total_runtime_ms":1802,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0289,"confidence_interval":0.9347,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.091,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1695,"p95_latency_ms":2072,"p99_latency_ms":2222,"time_to_first_token_ms":292,"total_runtime_ms":1883,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0106,"confidence_interval":0.9467,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.625,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1650,"p95_latency_ms":2017,"p99_latency_ms":2164,"time_to_first_token_ms":310,"total_runtime_ms":1834,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.027,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4948,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":987,"p95_latency_ms":1206,"p99_latency_ms":1294,"time_to_first_token_ms":266,"total_runtime_ms":1097,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.024,"confidence_interval":0.9437,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0405,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":972,"p95_latency_ms":1188,"p99_latency_ms":1275,"time_to_first_token_ms":239,"total_runtime_ms":1080,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0171,"confidence_interval":0.9475,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.524,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":999,"p95_latency_ms":1220,"p99_latency_ms":1309,"time_to_first_token_ms":267,"total_runtime_ms":1109,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0211,"confidence_interval":0.9443,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.979,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1001,"p95_latency_ms":1223,"p99_latency_ms":1312,"time_to_first_token_ms":244,"total_runtime_ms":1112,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0182,"confidence_interval":0.945,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.7009,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":961,"p95_latency_ms":1175,"p99_latency_ms":1260,"time_to_first_token_ms":257,"total_runtime_ms":1068,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0287,"confidence_interval":0.9396,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9701,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1043,"p95_latency_ms":1274,"p99_latency_ms":1367,"time_to_first_token_ms":236,"total_runtime_ms":1158,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.018,"confidence_interval":0.9455,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4276,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":994,"p95_latency_ms":1215,"p99_latency_ms":1304,"time_to_first_token_ms":234,"total_runtime_ms":1105,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0288,"confidence_interval":0.9439,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5993,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":976,"p95_latency_ms":1193,"p99_latency_ms":1280,"time_to_first_token_ms":242,"total_runtime_ms":1085,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.013,"confidence_interval":0.9396,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5798,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":980,"p95_latency_ms":1198,"p99_latency_ms":1285,"time_to_first_token_ms":262,"total_runtime_ms":1089,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0154,"confidence_interval":0.9301,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.943,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":984,"p95_latency_ms":1203,"p99_latency_ms":1291,"time_to_first_token_ms":235,"total_runtime_ms":1094,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0289,"confidence_interval":0.9317,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3166,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":732,"p95_latency_ms":895,"p99_latency_ms":960,"time_to_first_token_ms":159,"total_runtime_ms":813,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0109,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9749,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":739,"p95_latency_ms":903,"p99_latency_ms":969,"time_to_first_token_ms":189,"total_runtime_ms":821,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0149,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7551,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":784,"p95_latency_ms":958,"p99_latency_ms":1027,"time_to_first_token_ms":182,"total_runtime_ms":871,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0196,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7047,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":795,"p95_latency_ms":971,"p99_latency_ms":1042,"time_to_first_token_ms":173,"total_runtime_ms":883,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0274,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3066,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":788,"p95_latency_ms":963,"p99_latency_ms":1033,"time_to_first_token_ms":157,"total_runtime_ms":876,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0168,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2335,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":777,"p95_latency_ms":950,"p99_latency_ms":1019,"time_to_first_token_ms":161,"total_runtime_ms":863,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0201,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1063,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":732,"p95_latency_ms":895,"p99_latency_ms":960,"time_to_first_token_ms":191,"total_runtime_ms":813,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0103,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2993,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":768,"p95_latency_ms":938,"p99_latency_ms":1007,"time_to_first_token_ms":184,"total_runtime_ms":853,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0237,"confidence_interval":0.9452,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0741,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":728,"p95_latency_ms":889,"p99_latency_ms":954,"time_to_first_token_ms":179,"total_runtime_ms":809,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0129,"confidence_interval":0.9344,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6567,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":790,"p95_latency_ms":965,"p99_latency_ms":1035,"time_to_first_token_ms":186,"total_runtime_ms":878,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0205,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0507,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":630,"time_to_first_token_ms":133,"total_runtime_ms":534,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8575,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":465,"p95_latency_ms":569,"p99_latency_ms":610,"time_to_first_token_ms":141,"total_runtime_ms":517,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0181,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3975,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":132,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0155,"confidence_interval":0.9402,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3517,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":526,"p95_latency_ms":642,"p99_latency_ms":689,"time_to_first_token_ms":129,"total_runtime_ms":584,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8453,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":515,"p95_latency_ms":630,"p99_latency_ms":676,"time_to_first_token_ms":117,"total_runtime_ms":573,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0287,"confidence_interval":0.9448,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8722,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":495,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":123,"total_runtime_ms":550,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0228,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0234,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":466,"p95_latency_ms":570,"p99_latency_ms":611,"time_to_first_token_ms":111,"total_runtime_ms":518,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0297,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3699,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":428,"p95_latency_ms":524,"p99_latency_ms":562,"time_to_first_token_ms":143,"total_runtime_ms":476,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0108,"confidence_interval":0.9357,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5907,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":144,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0267,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6732,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":555,"time_to_first_token_ms":126,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0132,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.855,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1232,"p95_latency_ms":1506,"p99_latency_ms":1615,"time_to_first_token_ms":241,"total_runtime_ms":1369,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0269,"confidence_interval":0.9361,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4134,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1255,"p95_latency_ms":1534,"p99_latency_ms":1646,"time_to_first_token_ms":237,"total_runtime_ms":1395,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0269,"confidence_interval":0.9409,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9332,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1244,"p95_latency_ms":1521,"p99_latency_ms":1631,"time_to_first_token_ms":252,"total_runtime_ms":1383,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0279,"confidence_interval":0.9448,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2672,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1241,"p95_latency_ms":1517,"p99_latency_ms":1627,"time_to_first_token_ms":239,"total_runtime_ms":1379,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0247,"confidence_interval":0.9449,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6958,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1270,"p95_latency_ms":1552,"p99_latency_ms":1665,"time_to_first_token_ms":254,"total_runtime_ms":1411,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0285,"confidence_interval":0.9477,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5268,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1285,"p95_latency_ms":1570,"p99_latency_ms":1684,"time_to_first_token_ms":245,"total_runtime_ms":1427,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0271,"confidence_interval":0.9356,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1511,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1284,"p95_latency_ms":1570,"p99_latency_ms":1684,"time_to_first_token_ms":252,"total_runtime_ms":1427,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0158,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3799,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1249,"p95_latency_ms":1526,"p99_latency_ms":1637,"time_to_first_token_ms":251,"total_runtime_ms":1387,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0142,"confidence_interval":0.9482,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7764,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1282,"p95_latency_ms":1567,"p99_latency_ms":1681,"time_to_first_token_ms":238,"total_runtime_ms":1424,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0128,"confidence_interval":0.9305,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3721,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1207,"p95_latency_ms":1475,"p99_latency_ms":1583,"time_to_first_token_ms":257,"total_runtime_ms":1341,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0282,"confidence_interval":0.9473,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1521,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":780,"p95_latency_ms":954,"p99_latency_ms":1023,"time_to_first_token_ms":179,"total_runtime_ms":867,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.9321,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.031,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":766,"p95_latency_ms":936,"p99_latency_ms":1005,"time_to_first_token_ms":205,"total_runtime_ms":851,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0155,"confidence_interval":0.9486,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1925,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":763,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":210,"total_runtime_ms":848,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0283,"confidence_interval":0.9378,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7997,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":779,"p95_latency_ms":952,"p99_latency_ms":1022,"time_to_first_token_ms":184,"total_runtime_ms":866,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0159,"confidence_interval":0.9443,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3411,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":815,"p95_latency_ms":996,"p99_latency_ms":1069,"time_to_first_token_ms":191,"total_runtime_ms":906,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0157,"confidence_interval":0.9435,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7461,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":772,"p95_latency_ms":944,"p99_latency_ms":1012,"time_to_first_token_ms":196,"total_runtime_ms":858,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9399,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.124,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":740,"p95_latency_ms":905,"p99_latency_ms":971,"time_to_first_token_ms":203,"total_runtime_ms":823,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0183,"confidence_interval":0.9385,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0787,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":781,"p95_latency_ms":954,"p99_latency_ms":1023,"time_to_first_token_ms":188,"total_runtime_ms":867,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0259,"confidence_interval":0.9457,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2666,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":743,"p95_latency_ms":908,"p99_latency_ms":974,"time_to_first_token_ms":179,"total_runtime_ms":825,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0173,"confidence_interval":0.9318,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.942,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":731,"p95_latency_ms":894,"p99_latency_ms":959,"time_to_first_token_ms":205,"total_runtime_ms":813,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0253,"confidence_interval":0.9441,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8638,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":599,"p95_latency_ms":732,"p99_latency_ms":785,"time_to_first_token_ms":137,"total_runtime_ms":665,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0263,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4287,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":624,"p95_latency_ms":762,"p99_latency_ms":818,"time_to_first_token_ms":150,"total_runtime_ms":693,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0161,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3562,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":755,"time_to_first_token_ms":144,"total_runtime_ms":640,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0158,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.754,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":610,"p95_latency_ms":746,"p99_latency_ms":800,"time_to_first_token_ms":164,"total_runtime_ms":678,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0225,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2616,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":629,"p95_latency_ms":769,"p99_latency_ms":825,"time_to_first_token_ms":161,"total_runtime_ms":699,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0156,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5029,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":570,"p95_latency_ms":697,"p99_latency_ms":748,"time_to_first_token_ms":154,"total_runtime_ms":634,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0195,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7667,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":154,"total_runtime_ms":635,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0262,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1231,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":634,"p95_latency_ms":775,"p99_latency_ms":831,"time_to_first_token_ms":165,"total_runtime_ms":704,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0274,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8171,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":614,"p95_latency_ms":750,"p99_latency_ms":805,"time_to_first_token_ms":135,"total_runtime_ms":682,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0143,"confidence_interval":0.9331,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5118,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":532,"p95_latency_ms":651,"p99_latency_ms":698,"time_to_first_token_ms":157,"total_runtime_ms":591,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0216,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1516,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":351,"p95_latency_ms":428,"p99_latency_ms":460,"time_to_first_token_ms":101,"total_runtime_ms":389,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0172,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6961,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":419,"p95_latency_ms":512,"p99_latency_ms":549,"time_to_first_token_ms":88,"total_runtime_ms":465,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0114,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0078,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":495,"time_to_first_token_ms":106,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0142,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3951,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":555,"time_to_first_token_ms":107,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0224,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4522,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":477,"time_to_first_token_ms":104,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.011,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4824,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":411,"p95_latency_ms":502,"p99_latency_ms":539,"time_to_first_token_ms":92,"total_runtime_ms":457,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0241,"confidence_interval":0.9495,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0552,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":423,"p95_latency_ms":517,"p99_latency_ms":554,"time_to_first_token_ms":97,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0111,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0841,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":380,"p95_latency_ms":464,"p99_latency_ms":498,"time_to_first_token_ms":102,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0291,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2342,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":423,"p95_latency_ms":516,"p99_latency_ms":554,"time_to_first_token_ms":99,"total_runtime_ms":469,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0249,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4554,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":415,"p95_latency_ms":508,"p99_latency_ms":545,"time_to_first_token_ms":104,"total_runtime_ms":462,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0236,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1789,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":926,"p95_latency_ms":1132,"p99_latency_ms":1214,"time_to_first_token_ms":203,"total_runtime_ms":1029,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0205,"confidence_interval":0.9368,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.518,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":971,"p95_latency_ms":1186,"p99_latency_ms":1273,"time_to_first_token_ms":204,"total_runtime_ms":1079,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0133,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7307,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":968,"p95_latency_ms":1183,"p99_latency_ms":1269,"time_to_first_token_ms":190,"total_runtime_ms":1076,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0217,"confidence_interval":0.942,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1155,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":944,"p95_latency_ms":1154,"p99_latency_ms":1238,"time_to_first_token_ms":212,"total_runtime_ms":1049,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0103,"confidence_interval":0.9459,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2696,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":911,"p95_latency_ms":1114,"p99_latency_ms":1195,"time_to_first_token_ms":197,"total_runtime_ms":1013,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0237,"confidence_interval":0.9397,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2874,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":983,"p95_latency_ms":1201,"p99_latency_ms":1289,"time_to_first_token_ms":199,"total_runtime_ms":1092,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0293,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9891,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":962,"p95_latency_ms":1176,"p99_latency_ms":1262,"time_to_first_token_ms":189,"total_runtime_ms":1069,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0246,"confidence_interval":0.9445,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8159,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":936,"p95_latency_ms":1144,"p99_latency_ms":1227,"time_to_first_token_ms":201,"total_runtime_ms":1040,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0138,"confidence_interval":0.9347,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.716,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":996,"p95_latency_ms":1217,"p99_latency_ms":1306,"time_to_first_token_ms":196,"total_runtime_ms":1106,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0276,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0023,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":988,"p95_latency_ms":1208,"p99_latency_ms":1296,"time_to_first_token_ms":215,"total_runtime_ms":1098,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.012,"confidence_interval":0.9437,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8096,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":576,"p95_latency_ms":703,"p99_latency_ms":755,"time_to_first_token_ms":144,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0131,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2676,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":584,"p95_latency_ms":714,"p99_latency_ms":766,"time_to_first_token_ms":143,"total_runtime_ms":649,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0217,"confidence_interval":0.93,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6305,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":540,"p95_latency_ms":660,"p99_latency_ms":708,"time_to_first_token_ms":166,"total_runtime_ms":600,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0293,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1552,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":580,"p95_latency_ms":709,"p99_latency_ms":760,"time_to_first_token_ms":161,"total_runtime_ms":644,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0155,"confidence_interval":0.9365,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2101,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":142,"total_runtime_ms":635,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.023,"confidence_interval":0.9489,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4604,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":577,"p95_latency_ms":706,"p99_latency_ms":757,"time_to_first_token_ms":161,"total_runtime_ms":642,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0195,"confidence_interval":0.946,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8138,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":614,"p95_latency_ms":751,"p99_latency_ms":806,"time_to_first_token_ms":148,"total_runtime_ms":683,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0205,"confidence_interval":0.9345,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.418,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":757,"time_to_first_token_ms":163,"total_runtime_ms":641,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0127,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1578,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":567,"p95_latency_ms":693,"p99_latency_ms":743,"time_to_first_token_ms":142,"total_runtime_ms":630,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0152,"confidence_interval":0.9375,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7278,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":169,"total_runtime_ms":635,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0149,"confidence_interval":0.9469,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8492,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":601,"time_to_first_token_ms":153,"total_runtime_ms":510,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0134,"confidence_interval":0.9317,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1785,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":467,"p95_latency_ms":571,"p99_latency_ms":612,"time_to_first_token_ms":147,"total_runtime_ms":519,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0273,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9142,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":511,"p95_latency_ms":625,"p99_latency_ms":670,"time_to_first_token_ms":132,"total_runtime_ms":568,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0189,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4616,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":513,"p95_latency_ms":627,"p99_latency_ms":673,"time_to_first_token_ms":137,"total_runtime_ms":570,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0217,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5351,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":506,"p95_latency_ms":618,"p99_latency_ms":663,"time_to_first_token_ms":150,"total_runtime_ms":562,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0104,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.78,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":497,"p95_latency_ms":608,"p99_latency_ms":652,"time_to_first_token_ms":137,"total_runtime_ms":552,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0257,"confidence_interval":0.9434,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0145,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":469,"p95_latency_ms":573,"p99_latency_ms":615,"time_to_first_token_ms":133,"total_runtime_ms":521,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0137,"confidence_interval":0.9438,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1682,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":494,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":121,"total_runtime_ms":549,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0222,"confidence_interval":0.9374,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7674,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":509,"p95_latency_ms":622,"p99_latency_ms":667,"time_to_first_token_ms":120,"total_runtime_ms":565,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.026,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9949,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":465,"p95_latency_ms":568,"p99_latency_ms":610,"time_to_first_token_ms":126,"total_runtime_ms":517,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0169,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6165,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":278,"p95_latency_ms":340,"p99_latency_ms":365,"time_to_first_token_ms":107,"total_runtime_ms":309,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0123,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9675,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":341,"p95_latency_ms":417,"p99_latency_ms":447,"time_to_first_token_ms":75,"total_runtime_ms":379,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0136,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4372,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":292,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":98,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0288,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9438,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":438,"p99_latency_ms":470,"time_to_first_token_ms":79,"total_runtime_ms":398,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0139,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.686,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":286,"p95_latency_ms":350,"p99_latency_ms":375,"time_to_first_token_ms":102,"total_runtime_ms":318,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0199,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5177,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":360,"p95_latency_ms":440,"p99_latency_ms":472,"time_to_first_token_ms":91,"total_runtime_ms":400,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0163,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0605,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":438,"p99_latency_ms":470,"time_to_first_token_ms":109,"total_runtime_ms":398,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0133,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6612,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":360,"p95_latency_ms":440,"p99_latency_ms":472,"time_to_first_token_ms":79,"total_runtime_ms":400,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0233,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4393,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":323,"p95_latency_ms":394,"p99_latency_ms":423,"time_to_first_token_ms":100,"total_runtime_ms":359,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0276,"confidence_interval":0.9408,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0859,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":294,"p95_latency_ms":360,"p99_latency_ms":386,"time_to_first_token_ms":99,"total_runtime_ms":327,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.029,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8222,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":791,"p95_latency_ms":967,"p99_latency_ms":1037,"time_to_first_token_ms":181,"total_runtime_ms":879,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.026,"confidence_interval":0.9316,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.332,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":763,"p95_latency_ms":933,"p99_latency_ms":1001,"time_to_first_token_ms":196,"total_runtime_ms":848,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0107,"confidence_interval":0.9363,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.893,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":789,"p95_latency_ms":965,"p99_latency_ms":1035,"time_to_first_token_ms":183,"total_runtime_ms":877,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0258,"confidence_interval":0.9445,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1116,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":784,"p95_latency_ms":958,"p99_latency_ms":1028,"time_to_first_token_ms":164,"total_runtime_ms":871,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.9362,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8601,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":768,"p95_latency_ms":939,"p99_latency_ms":1007,"time_to_first_token_ms":173,"total_runtime_ms":854,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0214,"confidence_interval":0.9344,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4132,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":758,"p95_latency_ms":927,"p99_latency_ms":994,"time_to_first_token_ms":189,"total_runtime_ms":842,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0166,"confidence_interval":0.931,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1085,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":778,"p95_latency_ms":951,"p99_latency_ms":1020,"time_to_first_token_ms":185,"total_runtime_ms":865,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0132,"confidence_interval":0.9337,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8577,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":785,"p95_latency_ms":960,"p99_latency_ms":1030,"time_to_first_token_ms":175,"total_runtime_ms":873,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.02,"confidence_interval":0.9366,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.13,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":809,"p95_latency_ms":989,"p99_latency_ms":1061,"time_to_first_token_ms":193,"total_runtime_ms":899,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.01,"confidence_interval":0.9494,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5806,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":854,"p95_latency_ms":1043,"p99_latency_ms":1119,"time_to_first_token_ms":191,"total_runtime_ms":948,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0101,"confidence_interval":0.9379,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7705,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":479,"p95_latency_ms":586,"p99_latency_ms":629,"time_to_first_token_ms":142,"total_runtime_ms":533,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0299,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8922,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":580,"p99_latency_ms":622,"time_to_first_token_ms":140,"total_runtime_ms":527,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0456,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0221,"confidence_interval":0.9451,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.121,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":523,"p95_latency_ms":639,"p99_latency_ms":685,"time_to_first_token_ms":150,"total_runtime_ms":581,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0167,"confidence_interval":0.9447,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1859,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":452,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":125,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0248,"confidence_interval":0.934,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3933,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":500,"p95_latency_ms":611,"p99_latency_ms":656,"time_to_first_token_ms":146,"total_runtime_ms":556,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0222,"confidence_interval":0.9437,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7866,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":541,"p95_latency_ms":662,"p99_latency_ms":710,"time_to_first_token_ms":127,"total_runtime_ms":602,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0279,"confidence_interval":0.9393,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2166,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":452,"p95_latency_ms":552,"p99_latency_ms":593,"time_to_first_token_ms":149,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0174,"confidence_interval":0.941,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3172,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":475,"p95_latency_ms":580,"p99_latency_ms":623,"time_to_first_token_ms":131,"total_runtime_ms":528,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.025,"confidence_interval":0.9363,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.357,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":479,"p95_latency_ms":585,"p99_latency_ms":628,"time_to_first_token_ms":132,"total_runtime_ms":532,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0212,"confidence_interval":0.9331,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8218,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"analyze_reasoning","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":487,"p95_latency_ms":595,"p99_latency_ms":639,"time_to_first_token_ms":146,"total_runtime_ms":541,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0171,"confidence_interval":0.9355,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0955,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1037,"p95_latency_ms":1267,"p99_latency_ms":1359,"time_to_first_token_ms":195,"total_runtime_ms":1152,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.029,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9096,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":993,"p95_latency_ms":1214,"p99_latency_ms":1302,"time_to_first_token_ms":210,"total_runtime_ms":1103,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0251,"confidence_interval":0.9308,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9237,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":995,"p95_latency_ms":1216,"p99_latency_ms":1305,"time_to_first_token_ms":194,"total_runtime_ms":1106,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0221,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0346,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1036,"p95_latency_ms":1266,"p99_latency_ms":1358,"time_to_first_token_ms":201,"total_runtime_ms":1151,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0206,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7204,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1030,"p95_latency_ms":1259,"p99_latency_ms":1350,"time_to_first_token_ms":201,"total_runtime_ms":1144,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0195,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1571,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":992,"p95_latency_ms":1213,"p99_latency_ms":1301,"time_to_first_token_ms":210,"total_runtime_ms":1103,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0172,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3419,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":995,"p95_latency_ms":1216,"p99_latency_ms":1305,"time_to_first_token_ms":206,"total_runtime_ms":1106,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0214,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6846,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":969,"p95_latency_ms":1184,"p99_latency_ms":1270,"time_to_first_token_ms":209,"total_runtime_ms":1077,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.019,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5573,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1008,"p95_latency_ms":1232,"p99_latency_ms":1321,"time_to_first_token_ms":217,"total_runtime_ms":1120,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0152,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1957,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":996,"p95_latency_ms":1217,"p99_latency_ms":1306,"time_to_first_token_ms":195,"total_runtime_ms":1107,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0132,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7888,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":656,"p95_latency_ms":802,"p99_latency_ms":860,"time_to_first_token_ms":177,"total_runtime_ms":729,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0195,"confidence_interval":0.9448,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5304,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":609,"p95_latency_ms":744,"p99_latency_ms":798,"time_to_first_token_ms":166,"total_runtime_ms":676,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0171,"confidence_interval":0.9346,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4355,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":628,"p95_latency_ms":767,"p99_latency_ms":823,"time_to_first_token_ms":149,"total_runtime_ms":697,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0197,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6413,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":700,"p99_latency_ms":751,"time_to_first_token_ms":174,"total_runtime_ms":636,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0259,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.174,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":611,"p95_latency_ms":747,"p99_latency_ms":802,"time_to_first_token_ms":148,"total_runtime_ms":679,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0187,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2236,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":584,"p95_latency_ms":713,"p99_latency_ms":765,"time_to_first_token_ms":173,"total_runtime_ms":648,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0113,"confidence_interval":0.9474,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3125,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":664,"p95_latency_ms":811,"p99_latency_ms":870,"time_to_first_token_ms":145,"total_runtime_ms":738,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0124,"confidence_interval":0.9322,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4828,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":619,"p95_latency_ms":756,"p99_latency_ms":811,"time_to_first_token_ms":162,"total_runtime_ms":688,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6628,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":641,"p95_latency_ms":783,"p99_latency_ms":840,"time_to_first_token_ms":153,"total_runtime_ms":712,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0292,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0738,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":612,"p95_latency_ms":748,"p99_latency_ms":802,"time_to_first_token_ms":166,"total_runtime_ms":680,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0296,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6834,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1625,"p95_latency_ms":1986,"p99_latency_ms":2131,"time_to_first_token_ms":312,"total_runtime_ms":1806,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.014,"confidence_interval":0.9301,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2365,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1593,"p95_latency_ms":1947,"p99_latency_ms":2089,"time_to_first_token_ms":303,"total_runtime_ms":1770,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0239,"confidence_interval":0.9337,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0383,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1639,"p95_latency_ms":2004,"p99_latency_ms":2150,"time_to_first_token_ms":311,"total_runtime_ms":1822,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0301,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0105,"confidence_interval":0.9498,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5188,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1672,"p95_latency_ms":2043,"p99_latency_ms":2192,"time_to_first_token_ms":297,"total_runtime_ms":1858,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.936,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6677,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1633,"p95_latency_ms":1996,"p99_latency_ms":2141,"time_to_first_token_ms":283,"total_runtime_ms":1815,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0196,"confidence_interval":0.9341,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.613,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1688,"p95_latency_ms":2064,"p99_latency_ms":2214,"time_to_first_token_ms":295,"total_runtime_ms":1876,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0254,"confidence_interval":0.9379,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0231,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1633,"p95_latency_ms":1996,"p99_latency_ms":2142,"time_to_first_token_ms":307,"total_runtime_ms":1815,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0224,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0628,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1681,"p95_latency_ms":2054,"p99_latency_ms":2204,"time_to_first_token_ms":312,"total_runtime_ms":1868,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0218,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6556,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1619,"p95_latency_ms":1979,"p99_latency_ms":2123,"time_to_first_token_ms":303,"total_runtime_ms":1799,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0169,"confidence_interval":0.9321,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4785,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1663,"p95_latency_ms":2033,"p99_latency_ms":2181,"time_to_first_token_ms":279,"total_runtime_ms":1848,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":87,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.028,"confidence_interval":0.9331,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8907,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":958,"p95_latency_ms":1170,"p99_latency_ms":1255,"time_to_first_token_ms":266,"total_runtime_ms":1064,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0289,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8822,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1016,"p95_latency_ms":1241,"p99_latency_ms":1332,"time_to_first_token_ms":263,"total_runtime_ms":1129,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0285,"confidence_interval":0.9315,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7373,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":957,"p95_latency_ms":1170,"p99_latency_ms":1255,"time_to_first_token_ms":249,"total_runtime_ms":1063,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0174,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4256,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":970,"p95_latency_ms":1186,"p99_latency_ms":1272,"time_to_first_token_ms":236,"total_runtime_ms":1078,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0125,"confidence_interval":0.9329,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3402,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1033,"p95_latency_ms":1263,"p99_latency_ms":1355,"time_to_first_token_ms":264,"total_runtime_ms":1148,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0266,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.277,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":960,"p95_latency_ms":1173,"p99_latency_ms":1258,"time_to_first_token_ms":247,"total_runtime_ms":1066,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.016,"confidence_interval":0.9377,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8125,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1050,"p95_latency_ms":1284,"p99_latency_ms":1377,"time_to_first_token_ms":249,"total_runtime_ms":1167,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0257,"confidence_interval":0.9302,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0026,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":998,"p95_latency_ms":1220,"p99_latency_ms":1308,"time_to_first_token_ms":261,"total_runtime_ms":1109,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0157,"confidence_interval":0.9314,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8556,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":961,"p95_latency_ms":1174,"p99_latency_ms":1259,"time_to_first_token_ms":239,"total_runtime_ms":1067,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0219,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9056,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":958,"p95_latency_ms":1171,"p99_latency_ms":1256,"time_to_first_token_ms":249,"total_runtime_ms":1065,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":77,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0153,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3879,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":719,"p95_latency_ms":879,"p99_latency_ms":943,"time_to_first_token_ms":171,"total_runtime_ms":799,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0161,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.75,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":755,"p95_latency_ms":923,"p99_latency_ms":991,"time_to_first_token_ms":174,"total_runtime_ms":839,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0185,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2643,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":704,"p95_latency_ms":860,"p99_latency_ms":923,"time_to_first_token_ms":182,"total_runtime_ms":782,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0218,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2303,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":731,"p95_latency_ms":893,"p99_latency_ms":958,"time_to_first_token_ms":181,"total_runtime_ms":812,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.012,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7896,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":782,"p95_latency_ms":955,"p99_latency_ms":1025,"time_to_first_token_ms":170,"total_runtime_ms":869,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0228,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9267,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":800,"p95_latency_ms":978,"p99_latency_ms":1050,"time_to_first_token_ms":160,"total_runtime_ms":889,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0254,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7021,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":732,"p95_latency_ms":895,"p99_latency_ms":960,"time_to_first_token_ms":171,"total_runtime_ms":814,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0393,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.019,"confidence_interval":0.9329,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4338,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":810,"p95_latency_ms":990,"p99_latency_ms":1062,"time_to_first_token_ms":173,"total_runtime_ms":900,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0183,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1059,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":758,"p95_latency_ms":926,"p99_latency_ms":994,"time_to_first_token_ms":158,"total_runtime_ms":842,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0275,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8968,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":753,"p95_latency_ms":920,"p99_latency_ms":987,"time_to_first_token_ms":159,"total_runtime_ms":837,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0152,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.426,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":449,"p95_latency_ms":549,"p99_latency_ms":589,"time_to_first_token_ms":139,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0264,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.667,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":439,"p95_latency_ms":536,"p99_latency_ms":575,"time_to_first_token_ms":123,"total_runtime_ms":487,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0264,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5667,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":494,"p95_latency_ms":604,"p99_latency_ms":647,"time_to_first_token_ms":138,"total_runtime_ms":549,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0233,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4599,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":503,"p95_latency_ms":615,"p99_latency_ms":660,"time_to_first_token_ms":131,"total_runtime_ms":559,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0246,"confidence_interval":0.9331,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8991,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":449,"p95_latency_ms":549,"p99_latency_ms":589,"time_to_first_token_ms":123,"total_runtime_ms":499,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0248,"confidence_interval":0.9452,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5852,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":501,"p95_latency_ms":613,"p99_latency_ms":657,"time_to_first_token_ms":126,"total_runtime_ms":557,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0189,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.406,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":500,"p95_latency_ms":611,"p99_latency_ms":656,"time_to_first_token_ms":116,"total_runtime_ms":556,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0172,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3285,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":424,"p95_latency_ms":518,"p99_latency_ms":556,"time_to_first_token_ms":139,"total_runtime_ms":471,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0258,"confidence_interval":0.9467,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9921,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":508,"p95_latency_ms":620,"p99_latency_ms":666,"time_to_first_token_ms":122,"total_runtime_ms":564,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0242,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1486,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":433,"p95_latency_ms":529,"p99_latency_ms":567,"time_to_first_token_ms":135,"total_runtime_ms":481,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0152,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0784,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1238,"p95_latency_ms":1513,"p99_latency_ms":1624,"time_to_first_token_ms":225,"total_runtime_ms":1376,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0224,"confidence_interval":0.9301,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.467,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1226,"p95_latency_ms":1499,"p99_latency_ms":1608,"time_to_first_token_ms":231,"total_runtime_ms":1363,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.016,"confidence_interval":0.9377,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8344,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1275,"p95_latency_ms":1559,"p99_latency_ms":1672,"time_to_first_token_ms":240,"total_runtime_ms":1417,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0137,"confidence_interval":0.9357,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1547,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1260,"p95_latency_ms":1540,"p99_latency_ms":1652,"time_to_first_token_ms":247,"total_runtime_ms":1400,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0131,"confidence_interval":0.9431,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0879,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1271,"p95_latency_ms":1554,"p99_latency_ms":1667,"time_to_first_token_ms":229,"total_runtime_ms":1412,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0131,"confidence_interval":0.9316,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8074,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1198,"p95_latency_ms":1465,"p99_latency_ms":1571,"time_to_first_token_ms":227,"total_runtime_ms":1332,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.018,"confidence_interval":0.9304,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.15,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1252,"p95_latency_ms":1530,"p99_latency_ms":1641,"time_to_first_token_ms":233,"total_runtime_ms":1391,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0107,"confidence_interval":0.9356,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0953,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1247,"p95_latency_ms":1525,"p99_latency_ms":1635,"time_to_first_token_ms":228,"total_runtime_ms":1386,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0228,"confidence_interval":0.9471,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7281,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1236,"p95_latency_ms":1511,"p99_latency_ms":1621,"time_to_first_token_ms":244,"total_runtime_ms":1373,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.02,"confidence_interval":0.9445,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9094,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1242,"p95_latency_ms":1519,"p99_latency_ms":1629,"time_to_first_token_ms":258,"total_runtime_ms":1380,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":68,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0234,"confidence_interval":0.9328,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.635,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":775,"p95_latency_ms":948,"p99_latency_ms":1017,"time_to_first_token_ms":194,"total_runtime_ms":861,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0126,"confidence_interval":0.9352,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2805,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":747,"p95_latency_ms":912,"p99_latency_ms":979,"time_to_first_token_ms":192,"total_runtime_ms":830,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.9437,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1362,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":799,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":200,"total_runtime_ms":887,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0225,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4111,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":730,"p95_latency_ms":892,"p99_latency_ms":957,"time_to_first_token_ms":206,"total_runtime_ms":811,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0194,"confidence_interval":0.9373,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2517,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":757,"p95_latency_ms":926,"p99_latency_ms":993,"time_to_first_token_ms":181,"total_runtime_ms":842,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0133,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0756,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":717,"p95_latency_ms":877,"p99_latency_ms":940,"time_to_first_token_ms":198,"total_runtime_ms":797,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0139,"confidence_interval":0.9367,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1852,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":769,"p95_latency_ms":940,"p99_latency_ms":1008,"time_to_first_token_ms":191,"total_runtime_ms":854,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0158,"confidence_interval":0.9377,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4893,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":793,"p95_latency_ms":969,"p99_latency_ms":1040,"time_to_first_token_ms":194,"total_runtime_ms":881,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0149,"confidence_interval":0.9348,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.935,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":735,"p95_latency_ms":898,"p99_latency_ms":964,"time_to_first_token_ms":201,"total_runtime_ms":817,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0115,"confidence_interval":0.9306,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8227,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":753,"p95_latency_ms":921,"p99_latency_ms":988,"time_to_first_token_ms":205,"total_runtime_ms":837,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":58,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0286,"confidence_interval":0.9366,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3902,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":599,"p95_latency_ms":733,"p99_latency_ms":786,"time_to_first_token_ms":147,"total_runtime_ms":666,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0168,"confidence_interval":0.9493,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7443,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":140,"total_runtime_ms":650,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0131,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4784,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":554,"p95_latency_ms":677,"p99_latency_ms":726,"time_to_first_token_ms":156,"total_runtime_ms":615,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0229,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.008,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":534,"p95_latency_ms":653,"p99_latency_ms":700,"time_to_first_token_ms":139,"total_runtime_ms":593,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.011,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3796,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":772,"time_to_first_token_ms":152,"total_runtime_ms":654,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0183,"confidence_interval":0.9388,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2714,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":558,"p95_latency_ms":682,"p99_latency_ms":732,"time_to_first_token_ms":152,"total_runtime_ms":620,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0113,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.63,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":579,"p95_latency_ms":708,"p99_latency_ms":759,"time_to_first_token_ms":144,"total_runtime_ms":643,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0241,"confidence_interval":0.9306,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7986,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":155,"total_runtime_ms":650,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0146,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4597,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":545,"p95_latency_ms":666,"p99_latency_ms":714,"time_to_first_token_ms":153,"total_runtime_ms":605,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.026,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1129,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":589,"p95_latency_ms":720,"p99_latency_ms":773,"time_to_first_token_ms":135,"total_runtime_ms":655,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":30,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0147,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5899,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":477,"time_to_first_token_ms":110,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0151,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5129,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":415,"p95_latency_ms":507,"p99_latency_ms":543,"time_to_first_token_ms":87,"total_runtime_ms":461,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0233,"confidence_interval":0.9416,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5444,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":347,"p95_latency_ms":425,"p99_latency_ms":455,"time_to_first_token_ms":87,"total_runtime_ms":386,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0125,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7138,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":390,"p95_latency_ms":477,"p99_latency_ms":511,"time_to_first_token_ms":100,"total_runtime_ms":433,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0273,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8973,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":350,"p95_latency_ms":428,"p99_latency_ms":459,"time_to_first_token_ms":110,"total_runtime_ms":389,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0194,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8342,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":362,"p95_latency_ms":443,"p99_latency_ms":475,"time_to_first_token_ms":118,"total_runtime_ms":402,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0239,"confidence_interval":0.9305,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.526,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":403,"p95_latency_ms":492,"p99_latency_ms":528,"time_to_first_token_ms":95,"total_runtime_ms":448,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9327,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3634,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":345,"p95_latency_ms":421,"p99_latency_ms":452,"time_to_first_token_ms":93,"total_runtime_ms":383,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0226,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0903,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":383,"p95_latency_ms":468,"p99_latency_ms":502,"time_to_first_token_ms":115,"total_runtime_ms":426,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0183,"confidence_interval":0.9442,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1729,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":334,"p95_latency_ms":408,"p99_latency_ms":438,"time_to_first_token_ms":110,"total_runtime_ms":371,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":30,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0115,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5992,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":938,"p95_latency_ms":1146,"p99_latency_ms":1230,"time_to_first_token_ms":204,"total_runtime_ms":1042,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0104,"confidence_interval":0.942,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9381,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":898,"p95_latency_ms":1098,"p99_latency_ms":1178,"time_to_first_token_ms":190,"total_runtime_ms":998,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0238,"confidence_interval":0.9466,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2473,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":961,"p95_latency_ms":1174,"p99_latency_ms":1260,"time_to_first_token_ms":214,"total_runtime_ms":1068,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0174,"confidence_interval":0.9388,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2897,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":911,"p95_latency_ms":1113,"p99_latency_ms":1194,"time_to_first_token_ms":217,"total_runtime_ms":1012,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0187,"confidence_interval":0.9405,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0422,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":921,"p95_latency_ms":1126,"p99_latency_ms":1208,"time_to_first_token_ms":204,"total_runtime_ms":1024,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0224,"confidence_interval":0.9467,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8056,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":974,"p95_latency_ms":1191,"p99_latency_ms":1277,"time_to_first_token_ms":184,"total_runtime_ms":1082,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0233,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2024,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":979,"p95_latency_ms":1196,"p99_latency_ms":1283,"time_to_first_token_ms":183,"total_runtime_ms":1087,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0241,"confidence_interval":0.9309,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3547,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":948,"p95_latency_ms":1158,"p99_latency_ms":1243,"time_to_first_token_ms":207,"total_runtime_ms":1053,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.015,"confidence_interval":0.9456,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6635,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":948,"p95_latency_ms":1158,"p99_latency_ms":1243,"time_to_first_token_ms":191,"total_runtime_ms":1053,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0174,"confidence_interval":0.9339,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1028,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":926,"p95_latency_ms":1132,"p99_latency_ms":1215,"time_to_first_token_ms":190,"total_runtime_ms":1029,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":51,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0129,"confidence_interval":0.9348,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0394,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":590,"p95_latency_ms":721,"p99_latency_ms":774,"time_to_first_token_ms":165,"total_runtime_ms":656,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0132,"confidence_interval":0.9436,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9991,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":559,"p95_latency_ms":683,"p99_latency_ms":733,"time_to_first_token_ms":163,"total_runtime_ms":621,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0159,"confidence_interval":0.9389,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5752,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":584,"p95_latency_ms":714,"p99_latency_ms":766,"time_to_first_token_ms":162,"total_runtime_ms":649,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0153,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7861,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":548,"p95_latency_ms":669,"p99_latency_ms":718,"time_to_first_token_ms":157,"total_runtime_ms":609,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0148,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8123,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":622,"p95_latency_ms":760,"p99_latency_ms":815,"time_to_first_token_ms":156,"total_runtime_ms":691,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0162,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5334,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":565,"p95_latency_ms":691,"p99_latency_ms":741,"time_to_first_token_ms":164,"total_runtime_ms":628,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0149,"confidence_interval":0.9471,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.003,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":147,"total_runtime_ms":636,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0252,"confidence_interval":0.9499,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6076,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":632,"p95_latency_ms":773,"p99_latency_ms":829,"time_to_first_token_ms":172,"total_runtime_ms":702,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0136,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1894,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":541,"p95_latency_ms":661,"p99_latency_ms":709,"time_to_first_token_ms":153,"total_runtime_ms":601,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0247,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9564,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":602,"p95_latency_ms":736,"p99_latency_ms":789,"time_to_first_token_ms":141,"total_runtime_ms":669,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":51,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0189,"confidence_interval":0.9417,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7781,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":447,"p95_latency_ms":546,"p99_latency_ms":586,"time_to_first_token_ms":124,"total_runtime_ms":496,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0111,"confidence_interval":0.944,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5163,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":441,"p95_latency_ms":539,"p99_latency_ms":578,"time_to_first_token_ms":129,"total_runtime_ms":490,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0197,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4589,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":463,"p95_latency_ms":566,"p99_latency_ms":607,"time_to_first_token_ms":131,"total_runtime_ms":514,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.028,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2546,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":464,"p95_latency_ms":567,"p99_latency_ms":608,"time_to_first_token_ms":124,"total_runtime_ms":516,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0224,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.938,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":459,"p95_latency_ms":562,"p99_latency_ms":602,"time_to_first_token_ms":137,"total_runtime_ms":511,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.017,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1947,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":522,"p95_latency_ms":638,"p99_latency_ms":685,"time_to_first_token_ms":128,"total_runtime_ms":580,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0213,"confidence_interval":0.9424,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1922,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":467,"p95_latency_ms":570,"p99_latency_ms":612,"time_to_first_token_ms":150,"total_runtime_ms":519,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0161,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7454,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":473,"p95_latency_ms":578,"p99_latency_ms":620,"time_to_first_token_ms":149,"total_runtime_ms":526,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0225,"confidence_interval":0.9329,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4783,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":442,"p95_latency_ms":540,"p99_latency_ms":579,"time_to_first_token_ms":126,"total_runtime_ms":491,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0285,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.012,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":446,"p95_latency_ms":545,"p99_latency_ms":584,"time_to_first_token_ms":148,"total_runtime_ms":495,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0146,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4812,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":326,"p95_latency_ms":399,"p99_latency_ms":428,"time_to_first_token_ms":102,"total_runtime_ms":362,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0223,"confidence_interval":0.9451,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.38,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":483,"time_to_first_token_ms":98,"total_runtime_ms":409,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0281,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1228,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":280,"p95_latency_ms":342,"p99_latency_ms":367,"time_to_first_token_ms":87,"total_runtime_ms":311,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0145,"confidence_interval":0.9314,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8156,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":354,"p95_latency_ms":433,"p99_latency_ms":465,"time_to_first_token_ms":95,"total_runtime_ms":394,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.026,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7848,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":325,"p95_latency_ms":397,"p99_latency_ms":426,"time_to_first_token_ms":99,"total_runtime_ms":361,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0152,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0837,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":289,"p95_latency_ms":354,"p99_latency_ms":379,"time_to_first_token_ms":99,"total_runtime_ms":321,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0155,"confidence_interval":0.9481,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4666,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":349,"p95_latency_ms":426,"p99_latency_ms":457,"time_to_first_token_ms":99,"total_runtime_ms":388,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0264,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7188,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":317,"p95_latency_ms":387,"p99_latency_ms":416,"time_to_first_token_ms":81,"total_runtime_ms":352,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0186,"confidence_interval":0.9431,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6675,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":318,"p95_latency_ms":389,"p99_latency_ms":417,"time_to_first_token_ms":76,"total_runtime_ms":353,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0262,"confidence_interval":0.949,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5067,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":346,"p95_latency_ms":423,"p99_latency_ms":454,"time_to_first_token_ms":89,"total_runtime_ms":385,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0191,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5161,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":750,"p95_latency_ms":917,"p99_latency_ms":984,"time_to_first_token_ms":167,"total_runtime_ms":834,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0297,"confidence_interval":0.9364,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4951,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":850,"p95_latency_ms":1039,"p99_latency_ms":1115,"time_to_first_token_ms":194,"total_runtime_ms":945,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0387,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0276,"confidence_interval":0.9324,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6848,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":791,"p95_latency_ms":967,"p99_latency_ms":1037,"time_to_first_token_ms":188,"total_runtime_ms":879,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0196,"confidence_interval":0.9359,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8938,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":773,"p95_latency_ms":944,"p99_latency_ms":1013,"time_to_first_token_ms":167,"total_runtime_ms":859,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0209,"confidence_interval":0.9366,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.7098,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":771,"p95_latency_ms":943,"p99_latency_ms":1011,"time_to_first_token_ms":176,"total_runtime_ms":857,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0231,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7932,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":841,"p95_latency_ms":1028,"p99_latency_ms":1103,"time_to_first_token_ms":193,"total_runtime_ms":935,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0128,"confidence_interval":0.9474,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8982,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":808,"p95_latency_ms":987,"p99_latency_ms":1059,"time_to_first_token_ms":166,"total_runtime_ms":897,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0127,"confidence_interval":0.9469,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8393,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":758,"p95_latency_ms":927,"p99_latency_ms":994,"time_to_first_token_ms":176,"total_runtime_ms":843,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0132,"confidence_interval":0.9436,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1542,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":817,"p95_latency_ms":999,"p99_latency_ms":1071,"time_to_first_token_ms":177,"total_runtime_ms":908,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.029,"confidence_interval":0.9418,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0472,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":850,"p95_latency_ms":1039,"p99_latency_ms":1115,"time_to_first_token_ms":185,"total_runtime_ms":945,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":46,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0207,"confidence_interval":0.9359,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.7042,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":480,"p95_latency_ms":587,"p99_latency_ms":630,"time_to_first_token_ms":148,"total_runtime_ms":534,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9395,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8676,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":494,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":129,"total_runtime_ms":549,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0273,"confidence_interval":0.9393,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.167,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":467,"p95_latency_ms":571,"p99_latency_ms":612,"time_to_first_token_ms":123,"total_runtime_ms":519,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0219,"confidence_interval":0.947,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9174,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":452,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":138,"total_runtime_ms":503,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0277,"confidence_interval":0.9458,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5121,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":519,"p95_latency_ms":635,"p99_latency_ms":681,"time_to_first_token_ms":127,"total_runtime_ms":577,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0337,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0132,"confidence_interval":0.943,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5648,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":539,"p95_latency_ms":659,"p99_latency_ms":707,"time_to_first_token_ms":117,"total_runtime_ms":599,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0216,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2174,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":508,"p95_latency_ms":621,"p99_latency_ms":666,"time_to_first_token_ms":124,"total_runtime_ms":564,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0165,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5304,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":451,"p95_latency_ms":552,"p99_latency_ms":592,"time_to_first_token_ms":139,"total_runtime_ms":501,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0238,"confidence_interval":0.9368,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3349,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":543,"p95_latency_ms":663,"p99_latency_ms":712,"time_to_first_token_ms":119,"total_runtime_ms":603,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.013,"confidence_interval":0.9396,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0565,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"compare_structured|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"compare_structured","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":118,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":36,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0114,"confidence_interval":0.9436,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2538,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1015,"p99_latency_ms":1089,"time_to_first_token_ms":196,"total_runtime_ms":923,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0194,"confidence_interval":0.9408,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7723,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":803,"p95_latency_ms":981,"p99_latency_ms":1052,"time_to_first_token_ms":199,"total_runtime_ms":892,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0297,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5446,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":855,"p95_latency_ms":1045,"p99_latency_ms":1122,"time_to_first_token_ms":208,"total_runtime_ms":950,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0172,"confidence_interval":0.9475,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6835,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":755,"p95_latency_ms":923,"p99_latency_ms":990,"time_to_first_token_ms":219,"total_runtime_ms":839,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0208,"confidence_interval":0.9376,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2467,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":759,"p95_latency_ms":928,"p99_latency_ms":996,"time_to_first_token_ms":189,"total_runtime_ms":844,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0172,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.243,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":773,"p95_latency_ms":945,"p99_latency_ms":1014,"time_to_first_token_ms":192,"total_runtime_ms":859,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0292,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2137,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":815,"p95_latency_ms":996,"p99_latency_ms":1069,"time_to_first_token_ms":194,"total_runtime_ms":906,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0234,"confidence_interval":0.9417,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5892,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":843,"p95_latency_ms":1031,"p99_latency_ms":1106,"time_to_first_token_ms":206,"total_runtime_ms":937,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0231,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7765,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":767,"p95_latency_ms":938,"p99_latency_ms":1006,"time_to_first_token_ms":216,"total_runtime_ms":853,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0297,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.087,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":761,"p95_latency_ms":930,"p99_latency_ms":998,"time_to_first_token_ms":197,"total_runtime_ms":845,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0193,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9248,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":428,"p95_latency_ms":523,"p99_latency_ms":561,"time_to_first_token_ms":157,"total_runtime_ms":475,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0264,"confidence_interval":0.9445,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7448,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":468,"p95_latency_ms":572,"p99_latency_ms":614,"time_to_first_token_ms":153,"total_runtime_ms":520,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0145,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6785,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":395,"p95_latency_ms":483,"p99_latency_ms":519,"time_to_first_token_ms":152,"total_runtime_ms":439,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0242,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9419,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":387,"p95_latency_ms":474,"p99_latency_ms":508,"time_to_first_token_ms":172,"total_runtime_ms":431,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0267,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4373,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":377,"p95_latency_ms":461,"p99_latency_ms":495,"time_to_first_token_ms":158,"total_runtime_ms":419,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0166,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9604,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":396,"p95_latency_ms":484,"p99_latency_ms":519,"time_to_first_token_ms":148,"total_runtime_ms":440,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0252,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.416,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":548,"time_to_first_token_ms":155,"total_runtime_ms":464,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0242,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7024,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":431,"p95_latency_ms":527,"p99_latency_ms":565,"time_to_first_token_ms":154,"total_runtime_ms":479,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0192,"confidence_interval":0.9301,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7411,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":440,"p95_latency_ms":538,"p99_latency_ms":577,"time_to_first_token_ms":165,"total_runtime_ms":489,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.021,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3479,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":391,"p95_latency_ms":478,"p99_latency_ms":513,"time_to_first_token_ms":152,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0229,"confidence_interval":0.9364,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9299,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1314,"p95_latency_ms":1606,"p99_latency_ms":1723,"time_to_first_token_ms":278,"total_runtime_ms":1460,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0262,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6159,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1332,"p95_latency_ms":1628,"p99_latency_ms":1746,"time_to_first_token_ms":298,"total_runtime_ms":1480,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.029,"confidence_interval":0.9329,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5194,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1334,"p95_latency_ms":1630,"p99_latency_ms":1749,"time_to_first_token_ms":293,"total_runtime_ms":1482,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.01,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6167,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1334,"p95_latency_ms":1631,"p99_latency_ms":1749,"time_to_first_token_ms":286,"total_runtime_ms":1482,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0203,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.874,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1274,"p95_latency_ms":1557,"p99_latency_ms":1671,"time_to_first_token_ms":287,"total_runtime_ms":1416,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0173,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4971,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1283,"p95_latency_ms":1568,"p99_latency_ms":1682,"time_to_first_token_ms":294,"total_runtime_ms":1425,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0296,"confidence_interval":0.9329,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5619,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1375,"p95_latency_ms":1680,"p99_latency_ms":1802,"time_to_first_token_ms":278,"total_runtime_ms":1527,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.013,"confidence_interval":0.9415,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5387,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1329,"p95_latency_ms":1624,"p99_latency_ms":1742,"time_to_first_token_ms":302,"total_runtime_ms":1476,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0157,"confidence_interval":0.9337,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8301,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1379,"p95_latency_ms":1685,"p99_latency_ms":1808,"time_to_first_token_ms":281,"total_runtime_ms":1532,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0251,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0408,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1305,"p95_latency_ms":1595,"p99_latency_ms":1711,"time_to_first_token_ms":283,"total_runtime_ms":1450,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0257,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3154,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":644,"p95_latency_ms":787,"p99_latency_ms":844,"time_to_first_token_ms":250,"total_runtime_ms":715,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0192,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0501,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":712,"p95_latency_ms":871,"p99_latency_ms":934,"time_to_first_token_ms":233,"total_runtime_ms":792,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0281,"confidence_interval":0.9442,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8703,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":701,"p95_latency_ms":857,"p99_latency_ms":920,"time_to_first_token_ms":266,"total_runtime_ms":779,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0289,"confidence_interval":0.9473,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3293,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":652,"p95_latency_ms":796,"p99_latency_ms":854,"time_to_first_token_ms":247,"total_runtime_ms":724,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.011,"confidence_interval":0.9327,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4742,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":699,"p95_latency_ms":854,"p99_latency_ms":917,"time_to_first_token_ms":260,"total_runtime_ms":777,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0285,"confidence_interval":0.9385,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7711,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":742,"p95_latency_ms":907,"p99_latency_ms":973,"time_to_first_token_ms":250,"total_runtime_ms":825,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0153,"confidence_interval":0.934,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5724,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":679,"p95_latency_ms":830,"p99_latency_ms":890,"time_to_first_token_ms":247,"total_runtime_ms":754,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9362,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4936,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":696,"p95_latency_ms":850,"p99_latency_ms":912,"time_to_first_token_ms":254,"total_runtime_ms":773,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0132,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8879,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":677,"p95_latency_ms":828,"p99_latency_ms":888,"time_to_first_token_ms":251,"total_runtime_ms":752,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0156,"confidence_interval":0.9379,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1357,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":658,"p95_latency_ms":804,"p99_latency_ms":863,"time_to_first_token_ms":242,"total_runtime_ms":731,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0212,"confidence_interval":0.9438,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3758,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":875,"time_to_first_token_ms":186,"total_runtime_ms":741,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.027,"confidence_interval":0.9495,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7614,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":582,"p95_latency_ms":711,"p99_latency_ms":763,"time_to_first_token_ms":174,"total_runtime_ms":646,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0469,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0123,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3156,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":633,"p95_latency_ms":774,"p99_latency_ms":830,"time_to_first_token_ms":180,"total_runtime_ms":704,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0152,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1951,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":850,"time_to_first_token_ms":160,"total_runtime_ms":721,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.017,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0767,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":575,"p95_latency_ms":702,"p99_latency_ms":753,"time_to_first_token_ms":161,"total_runtime_ms":639,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0108,"confidence_interval":0.9331,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4944,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":625,"p95_latency_ms":764,"p99_latency_ms":819,"time_to_first_token_ms":167,"total_runtime_ms":694,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0213,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3422,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":583,"p95_latency_ms":712,"p99_latency_ms":764,"time_to_first_token_ms":190,"total_runtime_ms":647,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0181,"confidence_interval":0.9327,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6415,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":637,"p95_latency_ms":779,"p99_latency_ms":835,"time_to_first_token_ms":158,"total_runtime_ms":708,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0204,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3859,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":633,"p95_latency_ms":774,"p99_latency_ms":831,"time_to_first_token_ms":159,"total_runtime_ms":704,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.011,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2658,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":669,"p95_latency_ms":818,"p99_latency_ms":877,"time_to_first_token_ms":167,"total_runtime_ms":743,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0266,"confidence_interval":0.9368,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3115,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":376,"p95_latency_ms":460,"p99_latency_ms":493,"time_to_first_token_ms":124,"total_runtime_ms":418,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0135,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5653,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":396,"p99_latency_ms":425,"time_to_first_token_ms":132,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0183,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5463,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":291,"p95_latency_ms":356,"p99_latency_ms":382,"time_to_first_token_ms":146,"total_runtime_ms":324,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0244,"confidence_interval":0.938,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5677,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":328,"p95_latency_ms":400,"p99_latency_ms":430,"time_to_first_token_ms":123,"total_runtime_ms":364,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0126,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8455,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":382,"p95_latency_ms":467,"p99_latency_ms":501,"time_to_first_token_ms":144,"total_runtime_ms":425,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0126,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8618,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":298,"p95_latency_ms":364,"p99_latency_ms":390,"time_to_first_token_ms":115,"total_runtime_ms":331,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0275,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.349,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":370,"p95_latency_ms":452,"p99_latency_ms":485,"time_to_first_token_ms":122,"total_runtime_ms":411,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0274,"confidence_interval":0.9402,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0103,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":326,"p95_latency_ms":398,"p99_latency_ms":427,"time_to_first_token_ms":135,"total_runtime_ms":362,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0156,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7634,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":478,"time_to_first_token_ms":116,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.025,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5144,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":301,"p95_latency_ms":368,"p99_latency_ms":395,"time_to_first_token_ms":126,"total_runtime_ms":335,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0255,"confidence_interval":0.9482,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0546,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1057,"p95_latency_ms":1292,"p99_latency_ms":1386,"time_to_first_token_ms":236,"total_runtime_ms":1175,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0131,"confidence_interval":0.932,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8409,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":992,"p95_latency_ms":1212,"p99_latency_ms":1300,"time_to_first_token_ms":224,"total_runtime_ms":1102,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0164,"confidence_interval":0.9431,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0773,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":999,"p95_latency_ms":1221,"p99_latency_ms":1310,"time_to_first_token_ms":234,"total_runtime_ms":1110,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0234,"confidence_interval":0.9442,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7784,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":971,"p95_latency_ms":1187,"p99_latency_ms":1273,"time_to_first_token_ms":257,"total_runtime_ms":1079,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0216,"confidence_interval":0.9375,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5206,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1018,"p95_latency_ms":1245,"p99_latency_ms":1335,"time_to_first_token_ms":246,"total_runtime_ms":1132,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0187,"confidence_interval":0.9408,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8617,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":960,"p95_latency_ms":1173,"p99_latency_ms":1259,"time_to_first_token_ms":244,"total_runtime_ms":1067,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0131,"confidence_interval":0.9492,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5126,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1040,"p95_latency_ms":1271,"p99_latency_ms":1364,"time_to_first_token_ms":255,"total_runtime_ms":1156,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0288,"confidence_interval":0.9498,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1804,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1036,"p95_latency_ms":1266,"p99_latency_ms":1358,"time_to_first_token_ms":248,"total_runtime_ms":1151,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0183,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5335,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":998,"p95_latency_ms":1219,"p99_latency_ms":1308,"time_to_first_token_ms":239,"total_runtime_ms":1108,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0222,"confidence_interval":0.9448,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6305,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":962,"p95_latency_ms":1176,"p99_latency_ms":1261,"time_to_first_token_ms":224,"total_runtime_ms":1069,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0243,"confidence_interval":0.9422,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3469,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":494,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":182,"total_runtime_ms":549,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0193,"confidence_interval":0.9448,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.551,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":755,"time_to_first_token_ms":200,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0246,"confidence_interval":0.9372,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8414,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":563,"p95_latency_ms":688,"p99_latency_ms":738,"time_to_first_token_ms":193,"total_runtime_ms":626,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0108,"confidence_interval":0.9482,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2804,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":519,"p95_latency_ms":635,"p99_latency_ms":681,"time_to_first_token_ms":189,"total_runtime_ms":577,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0245,"confidence_interval":0.9457,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8473,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":485,"p95_latency_ms":592,"p99_latency_ms":635,"time_to_first_token_ms":192,"total_runtime_ms":538,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.021,"confidence_interval":0.9317,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5234,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":549,"p95_latency_ms":671,"p99_latency_ms":720,"time_to_first_token_ms":208,"total_runtime_ms":610,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0117,"confidence_interval":0.9414,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0669,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":551,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":206,"total_runtime_ms":613,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0268,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0828,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":517,"p95_latency_ms":632,"p99_latency_ms":678,"time_to_first_token_ms":198,"total_runtime_ms":575,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0271,"confidence_interval":0.9314,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9361,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":521,"p95_latency_ms":637,"p99_latency_ms":683,"time_to_first_token_ms":212,"total_runtime_ms":579,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0296,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8355,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":756,"time_to_first_token_ms":183,"total_runtime_ms":641,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0106,"confidence_interval":0.9368,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8104,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":520,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":140,"total_runtime_ms":578,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0138,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5272,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":426,"p95_latency_ms":521,"p99_latency_ms":559,"time_to_first_token_ms":141,"total_runtime_ms":474,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0203,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0201,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":505,"p95_latency_ms":618,"p99_latency_ms":663,"time_to_first_token_ms":151,"total_runtime_ms":562,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.012,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":483,"p95_latency_ms":590,"p99_latency_ms":633,"time_to_first_token_ms":150,"total_runtime_ms":537,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0202,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6532,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":433,"p95_latency_ms":530,"p99_latency_ms":568,"time_to_first_token_ms":150,"total_runtime_ms":481,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0231,"confidence_interval":0.9442,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0272,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":520,"p95_latency_ms":636,"p99_latency_ms":682,"time_to_first_token_ms":150,"total_runtime_ms":578,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0164,"confidence_interval":0.9317,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4451,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":493,"p95_latency_ms":603,"p99_latency_ms":647,"time_to_first_token_ms":152,"total_runtime_ms":548,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0126,"confidence_interval":0.936,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8823,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":524,"p95_latency_ms":641,"p99_latency_ms":688,"time_to_first_token_ms":165,"total_runtime_ms":583,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0191,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7433,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":462,"p95_latency_ms":565,"p99_latency_ms":606,"time_to_first_token_ms":165,"total_runtime_ms":514,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.013,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0885,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":444,"p95_latency_ms":543,"p99_latency_ms":582,"time_to_first_token_ms":145,"total_runtime_ms":494,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.031,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0279,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2402,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":223,"p95_latency_ms":273,"p99_latency_ms":293,"time_to_first_token_ms":104,"total_runtime_ms":248,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4849,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":314,"p95_latency_ms":383,"p99_latency_ms":411,"time_to_first_token_ms":108,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0156,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7142,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":305,"p95_latency_ms":373,"p99_latency_ms":400,"time_to_first_token_ms":107,"total_runtime_ms":339,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0233,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7128,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":230,"p95_latency_ms":281,"p99_latency_ms":301,"time_to_first_token_ms":117,"total_runtime_ms":255,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0214,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3807,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":262,"p95_latency_ms":320,"p99_latency_ms":344,"time_to_first_token_ms":107,"total_runtime_ms":291,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0205,"confidence_interval":0.9378,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2373,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":311,"p95_latency_ms":380,"p99_latency_ms":408,"time_to_first_token_ms":106,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0257,"confidence_interval":0.9489,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4859,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":211,"p95_latency_ms":258,"p99_latency_ms":277,"time_to_first_token_ms":96,"total_runtime_ms":235,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0142,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8972,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":251,"p95_latency_ms":306,"p99_latency_ms":329,"time_to_first_token_ms":95,"total_runtime_ms":278,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0267,"confidence_interval":0.9496,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5471,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":231,"p95_latency_ms":282,"p99_latency_ms":303,"time_to_first_token_ms":111,"total_runtime_ms":257,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0415,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.012,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.575,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":308,"p95_latency_ms":376,"p99_latency_ms":404,"time_to_first_token_ms":112,"total_runtime_ms":342,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0278,"confidence_interval":0.938,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.313,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":883,"p99_latency_ms":948,"time_to_first_token_ms":186,"total_runtime_ms":803,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0117,"confidence_interval":0.9352,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8258,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":761,"p95_latency_ms":930,"p99_latency_ms":997,"time_to_first_token_ms":217,"total_runtime_ms":845,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0275,"confidence_interval":0.9483,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6017,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":725,"p95_latency_ms":886,"p99_latency_ms":950,"time_to_first_token_ms":192,"total_runtime_ms":805,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0211,"confidence_interval":0.9411,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6388,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":758,"p95_latency_ms":926,"p99_latency_ms":993,"time_to_first_token_ms":194,"total_runtime_ms":842,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0236,"confidence_interval":0.936,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3055,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":732,"p95_latency_ms":895,"p99_latency_ms":960,"time_to_first_token_ms":202,"total_runtime_ms":813,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.023,"confidence_interval":0.9428,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8212,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":749,"p95_latency_ms":916,"p99_latency_ms":983,"time_to_first_token_ms":212,"total_runtime_ms":833,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.045,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9399,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7555,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":746,"p95_latency_ms":911,"p99_latency_ms":978,"time_to_first_token_ms":202,"total_runtime_ms":828,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0242,"confidence_interval":0.9422,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0464,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":720,"p95_latency_ms":880,"p99_latency_ms":944,"time_to_first_token_ms":205,"total_runtime_ms":800,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0266,"confidence_interval":0.9333,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1646,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":767,"p95_latency_ms":937,"p99_latency_ms":1006,"time_to_first_token_ms":216,"total_runtime_ms":852,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0231,"confidence_interval":0.9317,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7442,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":744,"p95_latency_ms":910,"p99_latency_ms":976,"time_to_first_token_ms":216,"total_runtime_ms":827,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0222,"confidence_interval":0.9338,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1021,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":409,"p95_latency_ms":499,"p99_latency_ms":536,"time_to_first_token_ms":142,"total_runtime_ms":454,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0215,"confidence_interval":0.9322,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9002,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":416,"p95_latency_ms":508,"p99_latency_ms":545,"time_to_first_token_ms":149,"total_runtime_ms":462,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0423,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0199,"confidence_interval":0.9413,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6761,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":440,"p95_latency_ms":537,"p99_latency_ms":576,"time_to_first_token_ms":169,"total_runtime_ms":488,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0237,"confidence_interval":0.9447,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.345,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":413,"p95_latency_ms":505,"p99_latency_ms":541,"time_to_first_token_ms":153,"total_runtime_ms":459,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0141,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5882,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":432,"p95_latency_ms":528,"p99_latency_ms":567,"time_to_first_token_ms":166,"total_runtime_ms":480,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0114,"confidence_interval":0.9472,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1999,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":403,"p95_latency_ms":492,"p99_latency_ms":528,"time_to_first_token_ms":143,"total_runtime_ms":447,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0467,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0169,"confidence_interval":0.9482,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8855,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":399,"p95_latency_ms":487,"p99_latency_ms":523,"time_to_first_token_ms":164,"total_runtime_ms":443,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0217,"confidence_interval":0.9438,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.868,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":380,"p95_latency_ms":464,"p99_latency_ms":498,"time_to_first_token_ms":159,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.036,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0123,"confidence_interval":0.9441,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0298,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":408,"p95_latency_ms":498,"p99_latency_ms":534,"time_to_first_token_ms":171,"total_runtime_ms":453,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0113,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0512,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":425,"p95_latency_ms":520,"p99_latency_ms":558,"time_to_first_token_ms":164,"total_runtime_ms":473,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.013,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4743,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":397,"p95_latency_ms":485,"p99_latency_ms":521,"time_to_first_token_ms":136,"total_runtime_ms":441,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0224,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0198,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":416,"p95_latency_ms":508,"p99_latency_ms":545,"time_to_first_token_ms":121,"total_runtime_ms":462,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0278,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.507,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":359,"p95_latency_ms":439,"p99_latency_ms":471,"time_to_first_token_ms":146,"total_runtime_ms":399,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0228,"confidence_interval":0.9373,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4783,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":362,"p95_latency_ms":442,"p99_latency_ms":474,"time_to_first_token_ms":137,"total_runtime_ms":402,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0232,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3906,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":430,"p95_latency_ms":526,"p99_latency_ms":564,"time_to_first_token_ms":151,"total_runtime_ms":478,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0176,"confidence_interval":0.9454,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3247,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":419,"p95_latency_ms":512,"p99_latency_ms":550,"time_to_first_token_ms":122,"total_runtime_ms":466,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.016,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9817,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":351,"p95_latency_ms":429,"p99_latency_ms":460,"time_to_first_token_ms":121,"total_runtime_ms":390,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0112,"confidence_interval":0.9443,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3457,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":412,"p95_latency_ms":503,"p99_latency_ms":540,"time_to_first_token_ms":121,"total_runtime_ms":458,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0168,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6291,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":395,"p95_latency_ms":483,"p99_latency_ms":518,"time_to_first_token_ms":137,"total_runtime_ms":439,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0127,"confidence_interval":0.9388,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7574,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":406,"p95_latency_ms":496,"p99_latency_ms":532,"time_to_first_token_ms":130,"total_runtime_ms":451,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0233,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2227,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":222,"p95_latency_ms":271,"p99_latency_ms":291,"time_to_first_token_ms":103,"total_runtime_ms":246,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0385,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0285,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4216,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":222,"p95_latency_ms":271,"p99_latency_ms":291,"time_to_first_token_ms":87,"total_runtime_ms":246,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0466,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0197,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0594,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":209,"p95_latency_ms":255,"p99_latency_ms":274,"time_to_first_token_ms":75,"total_runtime_ms":232,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0338,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0136,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6082,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":251,"p95_latency_ms":307,"p99_latency_ms":329,"time_to_first_token_ms":89,"total_runtime_ms":279,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0165,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.836,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":211,"p95_latency_ms":258,"p99_latency_ms":277,"time_to_first_token_ms":98,"total_runtime_ms":235,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0277,"confidence_interval":0.9362,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1059,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":213,"p95_latency_ms":260,"p99_latency_ms":279,"time_to_first_token_ms":106,"total_runtime_ms":236,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0265,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2605,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":222,"p95_latency_ms":271,"p99_latency_ms":291,"time_to_first_token_ms":88,"total_runtime_ms":247,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0195,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8999,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":263,"p95_latency_ms":322,"p99_latency_ms":345,"time_to_first_token_ms":103,"total_runtime_ms":293,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0173,"confidence_interval":0.9347,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1237,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":208,"p95_latency_ms":254,"p99_latency_ms":272,"time_to_first_token_ms":87,"total_runtime_ms":231,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0273,"confidence_interval":0.9359,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2286,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":255,"p95_latency_ms":311,"p99_latency_ms":334,"time_to_first_token_ms":86,"total_runtime_ms":283,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0285,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1169,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":654,"p95_latency_ms":799,"p99_latency_ms":857,"time_to_first_token_ms":178,"total_runtime_ms":726,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.937,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2546,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":627,"p95_latency_ms":766,"p99_latency_ms":822,"time_to_first_token_ms":173,"total_runtime_ms":696,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0233,"confidence_interval":0.9378,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4425,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":164,"total_runtime_ms":665,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0314,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0249,"confidence_interval":0.9357,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1092,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":652,"p95_latency_ms":797,"p99_latency_ms":855,"time_to_first_token_ms":187,"total_runtime_ms":725,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0188,"confidence_interval":0.9359,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1191,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":678,"p95_latency_ms":829,"p99_latency_ms":890,"time_to_first_token_ms":195,"total_runtime_ms":754,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0139,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9762,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":613,"p95_latency_ms":749,"p99_latency_ms":804,"time_to_first_token_ms":167,"total_runtime_ms":681,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0351,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0225,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.384,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":646,"p95_latency_ms":790,"p99_latency_ms":848,"time_to_first_token_ms":186,"total_runtime_ms":718,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0194,"confidence_interval":0.9409,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7823,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":652,"p95_latency_ms":797,"p99_latency_ms":855,"time_to_first_token_ms":177,"total_runtime_ms":724,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0178,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0996,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":616,"p95_latency_ms":752,"p99_latency_ms":807,"time_to_first_token_ms":170,"total_runtime_ms":684,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0298,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0821,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":643,"p95_latency_ms":786,"p99_latency_ms":843,"time_to_first_token_ms":189,"total_runtime_ms":714,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0161,"confidence_interval":0.9405,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7873,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":312,"p95_latency_ms":381,"p99_latency_ms":409,"time_to_first_token_ms":124,"total_runtime_ms":346,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0136,"confidence_interval":0.9354,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1032,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":324,"p95_latency_ms":395,"p99_latency_ms":424,"time_to_first_token_ms":130,"total_runtime_ms":360,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0105,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9214,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":477,"time_to_first_token_ms":126,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0123,"confidence_interval":0.9361,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6155,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":358,"p95_latency_ms":438,"p99_latency_ms":470,"time_to_first_token_ms":147,"total_runtime_ms":398,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0132,"confidence_interval":0.9393,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9148,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":321,"p95_latency_ms":393,"p99_latency_ms":421,"time_to_first_token_ms":129,"total_runtime_ms":357,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0223,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5845,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":368,"p95_latency_ms":450,"p99_latency_ms":483,"time_to_first_token_ms":117,"total_runtime_ms":409,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.026,"confidence_interval":0.93,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9361,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":303,"p95_latency_ms":370,"p99_latency_ms":397,"time_to_first_token_ms":135,"total_runtime_ms":337,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0154,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0071,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":342,"p95_latency_ms":418,"p99_latency_ms":449,"time_to_first_token_ms":152,"total_runtime_ms":380,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0132,"confidence_interval":0.9336,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4225,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":338,"p95_latency_ms":413,"p99_latency_ms":443,"time_to_first_token_ms":126,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0285,"confidence_interval":0.9351,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3896,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"ambiguous_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":350,"p95_latency_ms":427,"p99_latency_ms":458,"time_to_first_token_ms":119,"total_runtime_ms":388,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0266,"confidence_interval":0.9321,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3696,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":905,"p95_latency_ms":1106,"p99_latency_ms":1186,"time_to_first_token_ms":221,"total_runtime_ms":1005,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0275,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1763,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":946,"p95_latency_ms":1156,"p99_latency_ms":1240,"time_to_first_token_ms":201,"total_runtime_ms":1051,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0497,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0256,"confidence_interval":0.9481,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1856,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":915,"p95_latency_ms":1118,"p99_latency_ms":1200,"time_to_first_token_ms":204,"total_runtime_ms":1017,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0166,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1278,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":924,"p95_latency_ms":1129,"p99_latency_ms":1211,"time_to_first_token_ms":199,"total_runtime_ms":1026,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0102,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5016,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":898,"p95_latency_ms":1098,"p99_latency_ms":1177,"time_to_first_token_ms":188,"total_runtime_ms":998,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0106,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2218,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":944,"p95_latency_ms":1154,"p99_latency_ms":1238,"time_to_first_token_ms":209,"total_runtime_ms":1049,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0171,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7378,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":898,"p95_latency_ms":1098,"p99_latency_ms":1177,"time_to_first_token_ms":203,"total_runtime_ms":998,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0159,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8913,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":896,"p95_latency_ms":1095,"p99_latency_ms":1175,"time_to_first_token_ms":213,"total_runtime_ms":996,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0117,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5909,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":977,"p95_latency_ms":1194,"p99_latency_ms":1281,"time_to_first_token_ms":219,"total_runtime_ms":1085,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.017,"confidence_interval":0.9331,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3865,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":936,"p95_latency_ms":1144,"p99_latency_ms":1227,"time_to_first_token_ms":216,"total_runtime_ms":1040,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0349,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0121,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5087,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":515,"p95_latency_ms":629,"p99_latency_ms":675,"time_to_first_token_ms":149,"total_runtime_ms":572,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0297,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2025,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":156,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0159,"confidence_interval":0.9323,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":608,"p95_latency_ms":743,"p99_latency_ms":797,"time_to_first_token_ms":157,"total_runtime_ms":675,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0348,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0219,"confidence_interval":0.9378,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9425,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":563,"p95_latency_ms":688,"p99_latency_ms":738,"time_to_first_token_ms":164,"total_runtime_ms":625,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0277,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6591,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":573,"p95_latency_ms":700,"p99_latency_ms":751,"time_to_first_token_ms":178,"total_runtime_ms":637,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0275,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3302,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":177,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0213,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2796,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":507,"p95_latency_ms":619,"p99_latency_ms":664,"time_to_first_token_ms":170,"total_runtime_ms":563,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0171,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9831,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":591,"p95_latency_ms":723,"p99_latency_ms":775,"time_to_first_token_ms":173,"total_runtime_ms":657,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0258,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7318,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":526,"p95_latency_ms":643,"p99_latency_ms":690,"time_to_first_token_ms":172,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0218,"confidence_interval":0.94,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2751,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":585,"p95_latency_ms":715,"p99_latency_ms":767,"time_to_first_token_ms":178,"total_runtime_ms":650,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0104,"confidence_interval":0.9481,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8916,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1574,"p95_latency_ms":1923,"p99_latency_ms":2063,"time_to_first_token_ms":282,"total_runtime_ms":1748,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0123,"confidence_interval":0.9433,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4341,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1594,"p95_latency_ms":1948,"p99_latency_ms":2089,"time_to_first_token_ms":288,"total_runtime_ms":1771,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.013,"confidence_interval":0.9365,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6123,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1509,"p95_latency_ms":1844,"p99_latency_ms":1978,"time_to_first_token_ms":291,"total_runtime_ms":1677,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0267,"confidence_interval":0.936,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2301,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1524,"p95_latency_ms":1863,"p99_latency_ms":1998,"time_to_first_token_ms":279,"total_runtime_ms":1693,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0112,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3174,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1528,"p95_latency_ms":1867,"p99_latency_ms":2003,"time_to_first_token_ms":297,"total_runtime_ms":1698,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0153,"confidence_interval":0.9328,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3468,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1511,"p95_latency_ms":1846,"p99_latency_ms":1981,"time_to_first_token_ms":281,"total_runtime_ms":1678,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.035,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0108,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.442,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1578,"p95_latency_ms":1929,"p99_latency_ms":2069,"time_to_first_token_ms":283,"total_runtime_ms":1754,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0115,"confidence_interval":0.9489,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0678,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1584,"p95_latency_ms":1936,"p99_latency_ms":2076,"time_to_first_token_ms":283,"total_runtime_ms":1760,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0192,"confidence_interval":0.9402,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.452,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1505,"p95_latency_ms":1840,"p99_latency_ms":1974,"time_to_first_token_ms":295,"total_runtime_ms":1673,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0153,"confidence_interval":0.9321,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2452,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1562,"p95_latency_ms":1909,"p99_latency_ms":2048,"time_to_first_token_ms":293,"total_runtime_ms":1735,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.024,"confidence_interval":0.9365,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6595,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":951,"p95_latency_ms":1162,"p99_latency_ms":1246,"time_to_first_token_ms":255,"total_runtime_ms":1056,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.029,"confidence_interval":0.9487,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0397,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":919,"p95_latency_ms":1123,"p99_latency_ms":1205,"time_to_first_token_ms":265,"total_runtime_ms":1021,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0234,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8886,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":873,"p95_latency_ms":1067,"p99_latency_ms":1145,"time_to_first_token_ms":234,"total_runtime_ms":970,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0196,"confidence_interval":0.9412,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7426,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":914,"p95_latency_ms":1117,"p99_latency_ms":1198,"time_to_first_token_ms":264,"total_runtime_ms":1016,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0153,"confidence_interval":0.941,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9015,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":926,"p95_latency_ms":1132,"p99_latency_ms":1215,"time_to_first_token_ms":261,"total_runtime_ms":1029,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0215,"confidence_interval":0.9393,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9742,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":946,"p95_latency_ms":1157,"p99_latency_ms":1241,"time_to_first_token_ms":251,"total_runtime_ms":1051,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.042,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0228,"confidence_interval":0.9401,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8147,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":871,"p95_latency_ms":1064,"p99_latency_ms":1142,"time_to_first_token_ms":247,"total_runtime_ms":967,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0197,"confidence_interval":0.9437,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8593,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":899,"p95_latency_ms":1099,"p99_latency_ms":1179,"time_to_first_token_ms":251,"total_runtime_ms":999,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0206,"confidence_interval":0.9315,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4643,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":870,"p95_latency_ms":1064,"p99_latency_ms":1141,"time_to_first_token_ms":261,"total_runtime_ms":967,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0204,"confidence_interval":0.9474,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3336,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":921,"p95_latency_ms":1125,"p99_latency_ms":1207,"time_to_first_token_ms":241,"total_runtime_ms":1023,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0262,"confidence_interval":0.9342,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9808,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":696,"p95_latency_ms":851,"p99_latency_ms":912,"time_to_first_token_ms":158,"total_runtime_ms":773,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0166,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.541,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":716,"p95_latency_ms":875,"p99_latency_ms":938,"time_to_first_token_ms":173,"total_runtime_ms":795,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0255,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6043,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":663,"p95_latency_ms":810,"p99_latency_ms":869,"time_to_first_token_ms":176,"total_runtime_ms":737,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0282,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8411,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":709,"p95_latency_ms":866,"p99_latency_ms":929,"time_to_first_token_ms":183,"total_runtime_ms":788,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0495,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0106,"confidence_interval":0.9325,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8631,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":732,"p95_latency_ms":894,"p99_latency_ms":959,"time_to_first_token_ms":167,"total_runtime_ms":813,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0197,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5023,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":675,"p95_latency_ms":825,"p99_latency_ms":885,"time_to_first_token_ms":182,"total_runtime_ms":750,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0453,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0219,"confidence_interval":0.9388,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4713,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":716,"p95_latency_ms":875,"p99_latency_ms":938,"time_to_first_token_ms":166,"total_runtime_ms":795,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0404,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0171,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1215,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":717,"p95_latency_ms":876,"p99_latency_ms":940,"time_to_first_token_ms":188,"total_runtime_ms":796,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0206,"confidence_interval":0.95,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2571,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":660,"p95_latency_ms":807,"p99_latency_ms":865,"time_to_first_token_ms":183,"total_runtime_ms":733,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0244,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2979,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":750,"p95_latency_ms":917,"p99_latency_ms":983,"time_to_first_token_ms":161,"total_runtime_ms":833,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0261,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1187,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":455,"p95_latency_ms":556,"p99_latency_ms":597,"time_to_first_token_ms":141,"total_runtime_ms":506,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0134,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6637,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":407,"p95_latency_ms":497,"p99_latency_ms":533,"time_to_first_token_ms":136,"total_runtime_ms":452,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0219,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0125,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":431,"p95_latency_ms":527,"p99_latency_ms":565,"time_to_first_token_ms":131,"total_runtime_ms":479,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0271,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8705,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":467,"p95_latency_ms":571,"p99_latency_ms":612,"time_to_first_token_ms":120,"total_runtime_ms":519,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7931,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":441,"p95_latency_ms":539,"p99_latency_ms":578,"time_to_first_token_ms":126,"total_runtime_ms":490,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0199,"confidence_interval":0.9427,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9712,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":452,"p95_latency_ms":552,"p99_latency_ms":593,"time_to_first_token_ms":145,"total_runtime_ms":502,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0208,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1421,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":420,"p95_latency_ms":513,"p99_latency_ms":550,"time_to_first_token_ms":128,"total_runtime_ms":466,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0265,"confidence_interval":0.9409,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1309,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":421,"p95_latency_ms":515,"p99_latency_ms":553,"time_to_first_token_ms":111,"total_runtime_ms":468,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0224,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.595,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":450,"p95_latency_ms":549,"p99_latency_ms":589,"time_to_first_token_ms":122,"total_runtime_ms":500,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0202,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2019,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":469,"p95_latency_ms":573,"p99_latency_ms":615,"time_to_first_token_ms":125,"total_runtime_ms":521,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0263,"confidence_interval":0.9483,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4259,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1185,"p95_latency_ms":1448,"p99_latency_ms":1553,"time_to_first_token_ms":231,"total_runtime_ms":1316,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0334,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0131,"confidence_interval":0.9388,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.242,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1133,"p95_latency_ms":1385,"p99_latency_ms":1486,"time_to_first_token_ms":233,"total_runtime_ms":1259,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0124,"confidence_interval":0.9493,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8281,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1152,"p95_latency_ms":1408,"p99_latency_ms":1510,"time_to_first_token_ms":235,"total_runtime_ms":1280,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0445,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0131,"confidence_interval":0.9407,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8883,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1186,"p95_latency_ms":1449,"p99_latency_ms":1555,"time_to_first_token_ms":230,"total_runtime_ms":1318,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0294,"confidence_interval":0.9382,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7513,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1137,"p95_latency_ms":1390,"p99_latency_ms":1491,"time_to_first_token_ms":230,"total_runtime_ms":1264,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0186,"confidence_interval":0.9468,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8655,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1209,"p95_latency_ms":1478,"p99_latency_ms":1585,"time_to_first_token_ms":258,"total_runtime_ms":1344,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0237,"confidence_interval":0.9463,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7316,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1197,"p95_latency_ms":1463,"p99_latency_ms":1569,"time_to_first_token_ms":248,"total_runtime_ms":1330,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0429,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0134,"confidence_interval":0.9342,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.7194,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1189,"p95_latency_ms":1453,"p99_latency_ms":1559,"time_to_first_token_ms":229,"total_runtime_ms":1321,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0132,"confidence_interval":0.9384,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6568,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1168,"p95_latency_ms":1427,"p99_latency_ms":1531,"time_to_first_token_ms":236,"total_runtime_ms":1297,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0257,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7564,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1172,"p95_latency_ms":1432,"p99_latency_ms":1537,"time_to_first_token_ms":254,"total_runtime_ms":1302,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":62,"gpu_utilization_percent":0,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0111,"confidence_interval":0.941,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4005,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":675,"p95_latency_ms":825,"p99_latency_ms":885,"time_to_first_token_ms":212,"total_runtime_ms":750,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0154,"confidence_interval":0.9374,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2255,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":727,"p95_latency_ms":889,"p99_latency_ms":954,"time_to_first_token_ms":181,"total_runtime_ms":808,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0167,"confidence_interval":0.9391,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3325,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":737,"p95_latency_ms":901,"p99_latency_ms":967,"time_to_first_token_ms":187,"total_runtime_ms":819,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0135,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4493,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":677,"p95_latency_ms":827,"p99_latency_ms":887,"time_to_first_token_ms":204,"total_runtime_ms":752,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0136,"confidence_interval":0.9468,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9944,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":851,"time_to_first_token_ms":205,"total_runtime_ms":721,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0286,"confidence_interval":0.947,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3044,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":650,"p95_latency_ms":795,"p99_latency_ms":852,"time_to_first_token_ms":207,"total_runtime_ms":722,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0161,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2883,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":647,"p95_latency_ms":791,"p99_latency_ms":848,"time_to_first_token_ms":181,"total_runtime_ms":719,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0275,"confidence_interval":0.9394,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1917,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":670,"p95_latency_ms":819,"p99_latency_ms":878,"time_to_first_token_ms":189,"total_runtime_ms":744,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0147,"confidence_interval":0.9312,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.662,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":702,"p95_latency_ms":858,"p99_latency_ms":921,"time_to_first_token_ms":190,"total_runtime_ms":780,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0242,"confidence_interval":0.9388,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1423,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":661,"p95_latency_ms":808,"p99_latency_ms":867,"time_to_first_token_ms":193,"total_runtime_ms":735,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":52,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0268,"confidence_interval":0.9476,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9546,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":505,"p95_latency_ms":617,"p99_latency_ms":662,"time_to_first_token_ms":156,"total_runtime_ms":561,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0178,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3723,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":508,"p95_latency_ms":621,"p99_latency_ms":666,"time_to_first_token_ms":160,"total_runtime_ms":564,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0186,"confidence_interval":0.946,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0936,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":496,"p95_latency_ms":606,"p99_latency_ms":650,"time_to_first_token_ms":133,"total_runtime_ms":551,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0465,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.025,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5126,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":580,"p95_latency_ms":709,"p99_latency_ms":760,"time_to_first_token_ms":165,"total_runtime_ms":644,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.013,"confidence_interval":0.9345,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1804,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":523,"p95_latency_ms":639,"p99_latency_ms":686,"time_to_first_token_ms":166,"total_runtime_ms":581,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.025,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4051,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":529,"p95_latency_ms":647,"p99_latency_ms":694,"time_to_first_token_ms":152,"total_runtime_ms":588,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.047,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.025,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0426,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":600,"p95_latency_ms":733,"p99_latency_ms":786,"time_to_first_token_ms":147,"total_runtime_ms":666,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.01,"confidence_interval":0.9441,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.563,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":577,"p95_latency_ms":705,"p99_latency_ms":756,"time_to_first_token_ms":164,"total_runtime_ms":641,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0242,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6634,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":560,"p95_latency_ms":685,"p99_latency_ms":734,"time_to_first_token_ms":166,"total_runtime_ms":622,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0213,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9348,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":551,"p95_latency_ms":673,"p99_latency_ms":722,"time_to_first_token_ms":134,"total_runtime_ms":612,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":49,"gpu_utilization_percent":27,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0163,"confidence_interval":0.9302,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8322,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":312,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":103,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4487,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":355,"p95_latency_ms":434,"p99_latency_ms":465,"time_to_first_token_ms":97,"total_runtime_ms":394,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0182,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.098,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":355,"p95_latency_ms":434,"p99_latency_ms":465,"time_to_first_token_ms":113,"total_runtime_ms":394,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0225,"confidence_interval":0.9369,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1254,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":359,"p95_latency_ms":438,"p99_latency_ms":470,"time_to_first_token_ms":108,"total_runtime_ms":399,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0325,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0108,"confidence_interval":0.9369,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0259,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":314,"p95_latency_ms":384,"p99_latency_ms":412,"time_to_first_token_ms":96,"total_runtime_ms":349,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0297,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4711,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":87,"total_runtime_ms":435,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0113,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9225,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":378,"p95_latency_ms":462,"p99_latency_ms":496,"time_to_first_token_ms":96,"total_runtime_ms":420,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0125,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.165,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":332,"p95_latency_ms":406,"p99_latency_ms":436,"time_to_first_token_ms":90,"total_runtime_ms":369,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0247,"confidence_interval":0.935,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.445,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":407,"time_to_first_token_ms":97,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0378,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0171,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8919,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":296,"p95_latency_ms":361,"p99_latency_ms":387,"time_to_first_token_ms":119,"total_runtime_ms":328,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":27,"oscillation_rate":0.0352,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0214,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8566,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":873,"p95_latency_ms":1067,"p99_latency_ms":1145,"time_to_first_token_ms":190,"total_runtime_ms":970,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.03,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0106,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7679,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":872,"p95_latency_ms":1066,"p99_latency_ms":1144,"time_to_first_token_ms":199,"total_runtime_ms":969,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0169,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.265,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":853,"p95_latency_ms":1043,"p99_latency_ms":1118,"time_to_first_token_ms":211,"total_runtime_ms":948,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0183,"confidence_interval":0.9409,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2118,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":863,"p95_latency_ms":1055,"p99_latency_ms":1132,"time_to_first_token_ms":207,"total_runtime_ms":959,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0209,"confidence_interval":0.95,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4938,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":884,"p95_latency_ms":1080,"p99_latency_ms":1158,"time_to_first_token_ms":209,"total_runtime_ms":982,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0305,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0179,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6669,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":857,"p95_latency_ms":1048,"p99_latency_ms":1124,"time_to_first_token_ms":217,"total_runtime_ms":952,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0294,"confidence_interval":0.9447,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6964,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":871,"p95_latency_ms":1064,"p99_latency_ms":1141,"time_to_first_token_ms":200,"total_runtime_ms":967,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0482,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0105,"confidence_interval":0.9432,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.593,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":912,"p95_latency_ms":1115,"p99_latency_ms":1196,"time_to_first_token_ms":192,"total_runtime_ms":1014,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0135,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9818,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":909,"p95_latency_ms":1111,"p99_latency_ms":1192,"time_to_first_token_ms":208,"total_runtime_ms":1010,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0209,"confidence_interval":0.9312,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9961,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":851,"p95_latency_ms":1040,"p99_latency_ms":1116,"time_to_first_token_ms":203,"total_runtime_ms":946,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":49,"gpu_utilization_percent":46,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0104,"confidence_interval":0.9339,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1668,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":562,"p95_latency_ms":687,"p99_latency_ms":736,"time_to_first_token_ms":164,"total_runtime_ms":624,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0242,"confidence_interval":0.9473,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1717,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":584,"p95_latency_ms":713,"p99_latency_ms":765,"time_to_first_token_ms":168,"total_runtime_ms":648,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0135,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7773,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":560,"p95_latency_ms":685,"p99_latency_ms":735,"time_to_first_token_ms":140,"total_runtime_ms":623,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0499,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0183,"confidence_interval":0.9431,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3023,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":588,"p95_latency_ms":719,"p99_latency_ms":771,"time_to_first_token_ms":155,"total_runtime_ms":654,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0129,"confidence_interval":0.943,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.12,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":543,"p95_latency_ms":664,"p99_latency_ms":712,"time_to_first_token_ms":141,"total_runtime_ms":603,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0138,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.5235,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":538,"p95_latency_ms":658,"p99_latency_ms":706,"time_to_first_token_ms":140,"total_runtime_ms":598,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0185,"confidence_interval":0.9428,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.7011,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":584,"p95_latency_ms":714,"p99_latency_ms":765,"time_to_first_token_ms":167,"total_runtime_ms":649,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.039,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0218,"confidence_interval":0.9493,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7261,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":506,"p95_latency_ms":619,"p99_latency_ms":664,"time_to_first_token_ms":161,"total_runtime_ms":562,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0395,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0283,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1859,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":502,"p95_latency_ms":614,"p99_latency_ms":658,"time_to_first_token_ms":170,"total_runtime_ms":558,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0142,"confidence_interval":0.9303,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6885,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":540,"p95_latency_ms":659,"p99_latency_ms":707,"time_to_first_token_ms":144,"total_runtime_ms":599,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":46,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0204,"confidence_interval":0.9469,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0452,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":509,"p95_latency_ms":623,"p99_latency_ms":668,"time_to_first_token_ms":155,"total_runtime_ms":566,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0141,"confidence_interval":0.9413,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4059,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":485,"p95_latency_ms":593,"p99_latency_ms":636,"time_to_first_token_ms":125,"total_runtime_ms":539,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0148,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8066,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":465,"p95_latency_ms":569,"p99_latency_ms":610,"time_to_first_token_ms":121,"total_runtime_ms":517,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0148,"confidence_interval":0.9446,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.435,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":510,"p95_latency_ms":623,"p99_latency_ms":668,"time_to_first_token_ms":124,"total_runtime_ms":566,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0175,"confidence_interval":0.9385,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9792,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":456,"p95_latency_ms":557,"p99_latency_ms":597,"time_to_first_token_ms":136,"total_runtime_ms":506,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0277,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.36,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":495,"p95_latency_ms":605,"p99_latency_ms":649,"time_to_first_token_ms":148,"total_runtime_ms":550,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0255,"confidence_interval":0.9402,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6451,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":491,"p95_latency_ms":600,"p99_latency_ms":644,"time_to_first_token_ms":133,"total_runtime_ms":545,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0227,"confidence_interval":0.9324,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1408,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":427,"p95_latency_ms":522,"p99_latency_ms":560,"time_to_first_token_ms":127,"total_runtime_ms":474,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0179,"confidence_interval":0.9358,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0652,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":497,"p95_latency_ms":607,"p99_latency_ms":652,"time_to_first_token_ms":134,"total_runtime_ms":552,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.038,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0263,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4343,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":414,"p95_latency_ms":506,"p99_latency_ms":543,"time_to_first_token_ms":150,"total_runtime_ms":460,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0231,"confidence_interval":0.9423,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8922,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":332,"p95_latency_ms":406,"p99_latency_ms":435,"time_to_first_token_ms":76,"total_runtime_ms":369,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0268,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4549,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":296,"p95_latency_ms":362,"p99_latency_ms":388,"time_to_first_token_ms":104,"total_runtime_ms":329,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0176,"confidence_interval":0.9396,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4977,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":317,"p95_latency_ms":387,"p99_latency_ms":416,"time_to_first_token_ms":85,"total_runtime_ms":352,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0392,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0149,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1176,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":300,"p95_latency_ms":367,"p99_latency_ms":393,"time_to_first_token_ms":82,"total_runtime_ms":333,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0216,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2843,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":406,"time_to_first_token_ms":76,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0298,"confidence_interval":0.9465,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2141,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":246,"p95_latency_ms":301,"p99_latency_ms":323,"time_to_first_token_ms":104,"total_runtime_ms":274,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0357,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0129,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4908,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":319,"p95_latency_ms":390,"p99_latency_ms":418,"time_to_first_token_ms":98,"total_runtime_ms":355,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.022,"confidence_interval":0.9452,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9693,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":309,"p95_latency_ms":378,"p99_latency_ms":406,"time_to_first_token_ms":87,"total_runtime_ms":344,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0153,"confidence_interval":0.9337,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9924,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":325,"p95_latency_ms":397,"p99_latency_ms":426,"time_to_first_token_ms":85,"total_runtime_ms":361,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0242,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7526,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":342,"p95_latency_ms":417,"p99_latency_ms":448,"time_to_first_token_ms":108,"total_runtime_ms":380,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0195,"confidence_interval":0.9414,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.1594,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":723,"p95_latency_ms":883,"p99_latency_ms":948,"time_to_first_token_ms":193,"total_runtime_ms":803,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0187,"confidence_interval":0.9452,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.7024,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":733,"p95_latency_ms":896,"p99_latency_ms":961,"time_to_first_token_ms":168,"total_runtime_ms":815,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.029,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.7112,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":720,"p95_latency_ms":880,"p99_latency_ms":944,"time_to_first_token_ms":178,"total_runtime_ms":800,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0267,"confidence_interval":0.9421,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7408,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":758,"p95_latency_ms":926,"p99_latency_ms":994,"time_to_first_token_ms":178,"total_runtime_ms":842,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0394,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0242,"confidence_interval":0.9318,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3149,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":790,"p95_latency_ms":966,"p99_latency_ms":1036,"time_to_first_token_ms":168,"total_runtime_ms":878,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0343,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0249,"confidence_interval":0.936,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3964,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":711,"p95_latency_ms":869,"p99_latency_ms":932,"time_to_first_token_ms":191,"total_runtime_ms":790,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0311,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0281,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1366,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":735,"p95_latency_ms":898,"p99_latency_ms":963,"time_to_first_token_ms":196,"total_runtime_ms":816,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0201,"confidence_interval":0.9391,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6567,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":744,"p95_latency_ms":910,"p99_latency_ms":976,"time_to_first_token_ms":185,"total_runtime_ms":827,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0367,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0176,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1366,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":705,"p95_latency_ms":862,"p99_latency_ms":924,"time_to_first_token_ms":190,"total_runtime_ms":783,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0234,"confidence_interval":0.9467,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4341,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":729,"p95_latency_ms":891,"p99_latency_ms":956,"time_to_first_token_ms":193,"total_runtime_ms":810,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":42,"gpu_utilization_percent":0,"oscillation_rate":0.0421,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":0,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0214,"confidence_interval":0.9415,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8034,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":494,"p95_latency_ms":604,"p99_latency_ms":648,"time_to_first_token_ms":149,"total_runtime_ms":549,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.9403,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8993,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":632,"time_to_first_token_ms":126,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0134,"confidence_interval":0.9341,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9382,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":476,"p95_latency_ms":582,"p99_latency_ms":624,"time_to_first_token_ms":147,"total_runtime_ms":529,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0331,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0257,"confidence_interval":0.948,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4087,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":596,"time_to_first_token_ms":150,"total_runtime_ms":505,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0178,"confidence_interval":0.9342,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1992,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":431,"p95_latency_ms":527,"p99_latency_ms":565,"time_to_first_token_ms":117,"total_runtime_ms":479,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0239,"confidence_interval":0.9344,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1777,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":404,"p95_latency_ms":494,"p99_latency_ms":530,"time_to_first_token_ms":121,"total_runtime_ms":449,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0149,"confidence_interval":0.9404,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3288,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":462,"p95_latency_ms":565,"p99_latency_ms":606,"time_to_first_token_ms":135,"total_runtime_ms":514,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0491,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0149,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5214,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":454,"p95_latency_ms":555,"p99_latency_ms":596,"time_to_first_token_ms":149,"total_runtime_ms":505,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0255,"confidence_interval":0.9343,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2299,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":417,"p95_latency_ms":510,"p99_latency_ms":547,"time_to_first_token_ms":149,"total_runtime_ms":463,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0109,"confidence_interval":0.9315,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4253,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"contradictory_intent","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.09,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":484,"p95_latency_ms":591,"p99_latency_ms":634,"time_to_first_token_ms":134,"total_runtime_ms":538,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":32,"gpu_utilization_percent":0,"oscillation_rate":0.0345,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0238,"confidence_interval":0.9464,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.995,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1145,"p95_latency_ms":1400,"p99_latency_ms":1502,"time_to_first_token_ms":217,"total_runtime_ms":1272,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0103,"confidence_interval":0.9307,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3201,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1091,"p95_latency_ms":1333,"p99_latency_ms":1430,"time_to_first_token_ms":199,"total_runtime_ms":1212,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0163,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9771,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1095,"p95_latency_ms":1338,"p99_latency_ms":1435,"time_to_first_token_ms":203,"total_runtime_ms":1216,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0232,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8483,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1063,"p95_latency_ms":1299,"p99_latency_ms":1393,"time_to_first_token_ms":199,"total_runtime_ms":1181,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0198,"confidence_interval":0.9495,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.663,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1149,"p95_latency_ms":1405,"p99_latency_ms":1507,"time_to_first_token_ms":202,"total_runtime_ms":1277,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0102,"confidence_interval":0.9419,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4508,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1083,"p95_latency_ms":1324,"p99_latency_ms":1420,"time_to_first_token_ms":219,"total_runtime_ms":1204,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0226,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2817,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1090,"p95_latency_ms":1332,"p99_latency_ms":1429,"time_to_first_token_ms":221,"total_runtime_ms":1211,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0277,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7807,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1149,"p95_latency_ms":1405,"p99_latency_ms":1507,"time_to_first_token_ms":191,"total_runtime_ms":1277,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0154,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5818,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1107,"p95_latency_ms":1352,"p99_latency_ms":1451,"time_to_first_token_ms":201,"total_runtime_ms":1230,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0297,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4424,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1156,"p95_latency_ms":1413,"p99_latency_ms":1516,"time_to_first_token_ms":191,"total_runtime_ms":1285,"cold_start_penalty_ms":416,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0296,"confidence_interval":0.9356,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4835,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":723,"p95_latency_ms":884,"p99_latency_ms":948,"time_to_first_token_ms":158,"total_runtime_ms":803,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0157,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.8706,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":772,"p95_latency_ms":944,"p99_latency_ms":1012,"time_to_first_token_ms":147,"total_runtime_ms":858,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0227,"confidence_interval":0.9418,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9303,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":749,"p95_latency_ms":915,"p99_latency_ms":982,"time_to_first_token_ms":150,"total_runtime_ms":832,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0276,"confidence_interval":0.9357,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0129,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":699,"p95_latency_ms":854,"p99_latency_ms":916,"time_to_first_token_ms":162,"total_runtime_ms":777,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0427,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0281,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2903,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":752,"p95_latency_ms":920,"p99_latency_ms":986,"time_to_first_token_ms":174,"total_runtime_ms":836,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.029,"confidence_interval":0.9432,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.907,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":778,"p95_latency_ms":950,"p99_latency_ms":1020,"time_to_first_token_ms":143,"total_runtime_ms":864,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0173,"confidence_interval":0.9319,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9826,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":742,"p95_latency_ms":907,"p99_latency_ms":973,"time_to_first_token_ms":174,"total_runtime_ms":825,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0204,"confidence_interval":0.9353,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.597,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":733,"p95_latency_ms":896,"p99_latency_ms":961,"time_to_first_token_ms":159,"total_runtime_ms":815,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0474,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0223,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4406,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":700,"p95_latency_ms":856,"p99_latency_ms":918,"time_to_first_token_ms":154,"total_runtime_ms":778,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0126,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9135,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":688,"p95_latency_ms":841,"p99_latency_ms":902,"time_to_first_token_ms":154,"total_runtime_ms":765,"cold_start_penalty_ms":0,"peak_ram_mb":1210,"avg_ram_mb":806,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0441,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0155,"confidence_interval":0.9422,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5424,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1847,"p95_latency_ms":2258,"p99_latency_ms":2422,"time_to_first_token_ms":281,"total_runtime_ms":2053,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.9398,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0511,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1842,"p95_latency_ms":2251,"p99_latency_ms":2415,"time_to_first_token_ms":291,"total_runtime_ms":2047,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9988,"variance_across_runs":0.0178,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6925,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1830,"p95_latency_ms":2237,"p99_latency_ms":2399,"time_to_first_token_ms":296,"total_runtime_ms":2033,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9975,"variance_across_runs":0.0278,"confidence_interval":0.9458,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2913,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1868,"p95_latency_ms":2283,"p99_latency_ms":2449,"time_to_first_token_ms":281,"total_runtime_ms":2075,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9962,"variance_across_runs":0.0174,"confidence_interval":0.9382,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3521,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1807,"p95_latency_ms":2208,"p99_latency_ms":2369,"time_to_first_token_ms":281,"total_runtime_ms":2007,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.995,"variance_across_runs":0.0221,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.946,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1855,"p95_latency_ms":2267,"p99_latency_ms":2432,"time_to_first_token_ms":281,"total_runtime_ms":2061,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0321,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9938,"variance_across_runs":0.0158,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8096,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1848,"p95_latency_ms":2259,"p99_latency_ms":2423,"time_to_first_token_ms":302,"total_runtime_ms":2053,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.034,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9925,"variance_across_runs":0.0159,"confidence_interval":0.9368,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2041,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1854,"p95_latency_ms":2266,"p99_latency_ms":2431,"time_to_first_token_ms":288,"total_runtime_ms":2060,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9912,"variance_across_runs":0.0197,"confidence_interval":0.9445,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.7164,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1877,"p95_latency_ms":2294,"p99_latency_ms":2461,"time_to_first_token_ms":299,"total_runtime_ms":2085,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.99,"variance_across_runs":0.0232,"confidence_interval":0.9382,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7653,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1846,"p95_latency_ms":2256,"p99_latency_ms":2420,"time_to_first_token_ms":307,"total_runtime_ms":2051,"cold_start_penalty_ms":707,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":100,"gpu_utilization_percent":0,"oscillation_rate":0.0426,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.18,"determinism_score":0.9888,"variance_across_runs":0.0176,"confidence_interval":0.9408,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8994,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1179,"p95_latency_ms":1441,"p99_latency_ms":1546,"time_to_first_token_ms":237,"total_runtime_ms":1310,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0119,"confidence_interval":0.9403,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3635,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1162,"p95_latency_ms":1420,"p99_latency_ms":1523,"time_to_first_token_ms":241,"total_runtime_ms":1291,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0169,"confidence_interval":0.933,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4489,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1168,"p95_latency_ms":1428,"p99_latency_ms":1532,"time_to_first_token_ms":253,"total_runtime_ms":1298,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0226,"confidence_interval":0.9442,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5303,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1186,"p95_latency_ms":1449,"p99_latency_ms":1555,"time_to_first_token_ms":251,"total_runtime_ms":1317,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0356,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0221,"confidence_interval":0.9468,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6207,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1185,"p95_latency_ms":1448,"p99_latency_ms":1553,"time_to_first_token_ms":250,"total_runtime_ms":1316,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0154,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9714,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1201,"p95_latency_ms":1468,"p99_latency_ms":1575,"time_to_first_token_ms":253,"total_runtime_ms":1335,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0347,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0165,"confidence_interval":0.9355,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9089,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1195,"p95_latency_ms":1461,"p99_latency_ms":1567,"time_to_first_token_ms":244,"total_runtime_ms":1328,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0439,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0245,"confidence_interval":0.9434,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.0035,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1186,"p95_latency_ms":1450,"p99_latency_ms":1555,"time_to_first_token_ms":256,"total_runtime_ms":1318,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.024,"confidence_interval":0.9447,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.005,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1192,"p95_latency_ms":1457,"p99_latency_ms":1563,"time_to_first_token_ms":233,"total_runtime_ms":1325,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0485,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0237,"confidence_interval":0.9458,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4025,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":1164,"p95_latency_ms":1423,"p99_latency_ms":1526,"time_to_first_token_ms":263,"total_runtime_ms":1293,"cold_start_penalty_ms":0,"peak_ram_mb":2056,"avg_ram_mb":1371,"cpu_utilization_percent":92,"gpu_utilization_percent":0,"oscillation_rate":0.0424,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0128,"confidence_interval":0.9422,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4813,"memory_management_cost":12.84,"memory_pressure_events":1} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":870,"p95_latency_ms":1064,"p99_latency_ms":1141,"time_to_first_token_ms":185,"total_runtime_ms":967,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0263,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3928,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":818,"p95_latency_ms":999,"p99_latency_ms":1072,"time_to_first_token_ms":182,"total_runtime_ms":909,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0288,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6811,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":826,"p95_latency_ms":1009,"p99_latency_ms":1083,"time_to_first_token_ms":174,"total_runtime_ms":918,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0354,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0285,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4874,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":830,"p95_latency_ms":1015,"p99_latency_ms":1089,"time_to_first_token_ms":157,"total_runtime_ms":922,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0257,"confidence_interval":0.9498,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.797,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":888,"p95_latency_ms":1085,"p99_latency_ms":1164,"time_to_first_token_ms":179,"total_runtime_ms":986,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.013,"confidence_interval":0.9459,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0413,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":799,"p95_latency_ms":976,"p99_latency_ms":1047,"time_to_first_token_ms":187,"total_runtime_ms":887,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0231,"confidence_interval":0.9352,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9854,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":810,"p95_latency_ms":991,"p99_latency_ms":1063,"time_to_first_token_ms":172,"total_runtime_ms":901,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0376,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0115,"confidence_interval":0.9433,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.1128,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":877,"p95_latency_ms":1072,"p99_latency_ms":1150,"time_to_first_token_ms":158,"total_runtime_ms":974,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0411,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0254,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6874,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":879,"p95_latency_ms":1074,"p99_latency_ms":1153,"time_to_first_token_ms":187,"total_runtime_ms":977,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.026,"confidence_interval":0.9372,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8594,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":836,"p95_latency_ms":1021,"p99_latency_ms":1096,"time_to_first_token_ms":187,"total_runtime_ms":929,"cold_start_penalty_ms":312,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0245,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.8243,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":516,"p95_latency_ms":631,"p99_latency_ms":677,"time_to_first_token_ms":142,"total_runtime_ms":574,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0279,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4183,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":610,"p95_latency_ms":746,"p99_latency_ms":800,"time_to_first_token_ms":136,"total_runtime_ms":678,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0223,"confidence_interval":0.9321,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.512,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":518,"p95_latency_ms":633,"p99_latency_ms":679,"time_to_first_token_ms":124,"total_runtime_ms":576,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0266,"confidence_interval":0.9375,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.765,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":542,"p95_latency_ms":662,"p99_latency_ms":711,"time_to_first_token_ms":121,"total_runtime_ms":602,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0217,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7697,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":561,"p95_latency_ms":685,"p99_latency_ms":735,"time_to_first_token_ms":145,"total_runtime_ms":623,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0187,"confidence_interval":0.9398,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6028,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":516,"p95_latency_ms":631,"p99_latency_ms":677,"time_to_first_token_ms":112,"total_runtime_ms":573,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0213,"confidence_interval":0.9321,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2853,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":671,"time_to_first_token_ms":124,"total_runtime_ms":569,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0455,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0167,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6146,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":523,"p95_latency_ms":639,"p99_latency_ms":686,"time_to_first_token_ms":138,"total_runtime_ms":581,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0132,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9887,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":527,"p95_latency_ms":644,"p99_latency_ms":691,"time_to_first_token_ms":130,"total_runtime_ms":585,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0431,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0133,"confidence_interval":0.9389,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.421,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":572,"p95_latency_ms":699,"p99_latency_ms":750,"time_to_first_token_ms":142,"total_runtime_ms":635,"cold_start_penalty_ms":0,"peak_ram_mb":907,"avg_ram_mb":605,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0408,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0251,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.004,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1384,"p95_latency_ms":1691,"p99_latency_ms":1814,"time_to_first_token_ms":248,"total_runtime_ms":1538,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0402,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0109,"confidence_interval":0.9398,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0462,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1388,"p95_latency_ms":1696,"p99_latency_ms":1819,"time_to_first_token_ms":249,"total_runtime_ms":1542,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0104,"confidence_interval":0.9314,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4129,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1427,"p95_latency_ms":1744,"p99_latency_ms":1871,"time_to_first_token_ms":234,"total_runtime_ms":1585,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0142,"confidence_interval":0.9422,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0046,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1355,"p95_latency_ms":1656,"p99_latency_ms":1776,"time_to_first_token_ms":235,"total_runtime_ms":1505,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0212,"confidence_interval":0.9407,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6736,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1372,"p95_latency_ms":1677,"p99_latency_ms":1799,"time_to_first_token_ms":235,"total_runtime_ms":1524,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0182,"confidence_interval":0.9316,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1354,"p95_latency_ms":1655,"p99_latency_ms":1775,"time_to_first_token_ms":245,"total_runtime_ms":1505,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0123,"confidence_interval":0.9424,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9352,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1429,"p95_latency_ms":1747,"p99_latency_ms":1874,"time_to_first_token_ms":254,"total_runtime_ms":1588,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0255,"confidence_interval":0.9316,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9389,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1421,"p95_latency_ms":1737,"p99_latency_ms":1863,"time_to_first_token_ms":250,"total_runtime_ms":1579,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0368,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0132,"confidence_interval":0.9461,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5921,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1389,"p95_latency_ms":1697,"p99_latency_ms":1821,"time_to_first_token_ms":257,"total_runtime_ms":1543,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.0332,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0117,"confidence_interval":0.9326,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5191,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1376,"p95_latency_ms":1682,"p99_latency_ms":1804,"time_to_first_token_ms":229,"total_runtime_ms":1529,"cold_start_penalty_ms":530,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":79,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0274,"confidence_interval":0.9395,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7651,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":884,"p95_latency_ms":1081,"p99_latency_ms":1160,"time_to_first_token_ms":204,"total_runtime_ms":983,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0451,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0177,"confidence_interval":0.9317,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4957,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":902,"p95_latency_ms":1102,"p99_latency_ms":1182,"time_to_first_token_ms":198,"total_runtime_ms":1002,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0181,"confidence_interval":0.9349,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.201,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":912,"p95_latency_ms":1114,"p99_latency_ms":1196,"time_to_first_token_ms":190,"total_runtime_ms":1013,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0234,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7956,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":875,"p95_latency_ms":1070,"p99_latency_ms":1148,"time_to_first_token_ms":184,"total_runtime_ms":973,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0177,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6334,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":866,"p95_latency_ms":1059,"p99_latency_ms":1136,"time_to_first_token_ms":195,"total_runtime_ms":962,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0269,"confidence_interval":0.9394,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7987,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":866,"p95_latency_ms":1059,"p99_latency_ms":1136,"time_to_first_token_ms":206,"total_runtime_ms":963,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0219,"confidence_interval":0.9373,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0897,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":936,"p95_latency_ms":1144,"p99_latency_ms":1227,"time_to_first_token_ms":187,"total_runtime_ms":1040,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0113,"confidence_interval":0.9479,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.9586,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":962,"p95_latency_ms":1175,"p99_latency_ms":1261,"time_to_first_token_ms":201,"total_runtime_ms":1068,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0246,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4166,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":904,"p95_latency_ms":1105,"p99_latency_ms":1186,"time_to_first_token_ms":195,"total_runtime_ms":1005,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.942,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.04,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":911,"p95_latency_ms":1113,"p99_latency_ms":1194,"time_to_first_token_ms":189,"total_runtime_ms":1012,"cold_start_penalty_ms":0,"peak_ram_mb":1542,"avg_ram_mb":1028,"cpu_utilization_percent":69,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0177,"confidence_interval":0.9448,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0142,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":637,"p95_latency_ms":779,"p99_latency_ms":836,"time_to_first_token_ms":140,"total_runtime_ms":708,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0401,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0292,"confidence_interval":0.9486,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0798,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":153,"total_runtime_ms":741,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0263,"confidence_interval":0.9478,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3661,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":613,"p95_latency_ms":749,"p99_latency_ms":803,"time_to_first_token_ms":142,"total_runtime_ms":681,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0298,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.322,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":628,"p95_latency_ms":767,"p99_latency_ms":823,"time_to_first_token_ms":157,"total_runtime_ms":697,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0479,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0281,"confidence_interval":0.9378,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0137,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":619,"p95_latency_ms":756,"p99_latency_ms":811,"time_to_first_token_ms":156,"total_runtime_ms":687,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0266,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6212,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":672,"p95_latency_ms":821,"p99_latency_ms":881,"time_to_first_token_ms":165,"total_runtime_ms":747,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0226,"confidence_interval":0.9367,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3598,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":613,"p95_latency_ms":750,"p99_latency_ms":804,"time_to_first_token_ms":163,"total_runtime_ms":681,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0473,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.019,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1084,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":632,"p95_latency_ms":773,"p99_latency_ms":829,"time_to_first_token_ms":135,"total_runtime_ms":702,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0282,"confidence_interval":0.9394,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6513,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":659,"p95_latency_ms":806,"p99_latency_ms":864,"time_to_first_token_ms":161,"total_runtime_ms":733,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0333,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0239,"confidence_interval":0.937,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5894,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":623,"p95_latency_ms":761,"p99_latency_ms":816,"time_to_first_token_ms":158,"total_runtime_ms":692,"cold_start_penalty_ms":234,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":62,"gpu_utilization_percent":36,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0164,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2125,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":394,"p95_latency_ms":481,"p99_latency_ms":517,"time_to_first_token_ms":114,"total_runtime_ms":438,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0327,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0123,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2638,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":406,"p95_latency_ms":496,"p99_latency_ms":532,"time_to_first_token_ms":111,"total_runtime_ms":451,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0442,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0133,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6635,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":482,"p95_latency_ms":589,"p99_latency_ms":631,"time_to_first_token_ms":104,"total_runtime_ms":535,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0407,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0219,"confidence_interval":0.9412,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5939,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":390,"p95_latency_ms":477,"p99_latency_ms":511,"time_to_first_token_ms":98,"total_runtime_ms":433,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0288,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8101,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":392,"p95_latency_ms":479,"p99_latency_ms":514,"time_to_first_token_ms":109,"total_runtime_ms":436,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0364,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0183,"confidence_interval":0.9361,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0673,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":396,"p95_latency_ms":484,"p99_latency_ms":519,"time_to_first_token_ms":105,"total_runtime_ms":440,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0241,"confidence_interval":0.9461,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3915,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":476,"p95_latency_ms":581,"p99_latency_ms":624,"time_to_first_token_ms":121,"total_runtime_ms":528,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0475,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0178,"confidence_interval":0.9471,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8386,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":476,"p95_latency_ms":582,"p99_latency_ms":624,"time_to_first_token_ms":89,"total_runtime_ms":529,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0123,"confidence_interval":0.9343,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1041,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":412,"p95_latency_ms":504,"p99_latency_ms":541,"time_to_first_token_ms":117,"total_runtime_ms":458,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0319,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0294,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0805,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":468,"p95_latency_ms":573,"p99_latency_ms":614,"time_to_first_token_ms":110,"total_runtime_ms":520,"cold_start_penalty_ms":0,"peak_ram_mb":680,"avg_ram_mb":454,"cpu_utilization_percent":52,"gpu_utilization_percent":36,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0273,"confidence_interval":0.9424,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.1598,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1083,"p95_latency_ms":1324,"p99_latency_ms":1420,"time_to_first_token_ms":207,"total_runtime_ms":1204,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.012,"confidence_interval":0.9389,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3267,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1043,"p95_latency_ms":1274,"p99_latency_ms":1367,"time_to_first_token_ms":211,"total_runtime_ms":1158,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.032,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.011,"confidence_interval":0.9402,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2399,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1024,"p95_latency_ms":1252,"p99_latency_ms":1343,"time_to_first_token_ms":205,"total_runtime_ms":1138,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0111,"confidence_interval":0.9406,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0021,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1106,"p95_latency_ms":1352,"p99_latency_ms":1450,"time_to_first_token_ms":188,"total_runtime_ms":1229,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0243,"confidence_interval":0.9426,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.304,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1018,"p95_latency_ms":1244,"p99_latency_ms":1335,"time_to_first_token_ms":211,"total_runtime_ms":1131,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0115,"confidence_interval":0.9368,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4692,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1038,"p95_latency_ms":1269,"p99_latency_ms":1361,"time_to_first_token_ms":216,"total_runtime_ms":1154,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0244,"confidence_interval":0.9372,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7564,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1040,"p95_latency_ms":1271,"p99_latency_ms":1364,"time_to_first_token_ms":193,"total_runtime_ms":1156,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.024,"confidence_interval":0.9373,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8403,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1009,"p95_latency_ms":1233,"p99_latency_ms":1323,"time_to_first_token_ms":209,"total_runtime_ms":1121,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0298,"confidence_interval":0.9421,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9695,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1096,"p95_latency_ms":1339,"p99_latency_ms":1437,"time_to_first_token_ms":213,"total_runtime_ms":1218,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0241,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2518,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":1103,"p95_latency_ms":1348,"p99_latency_ms":1446,"time_to_first_token_ms":189,"total_runtime_ms":1225,"cold_start_penalty_ms":398,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":62,"gpu_utilization_percent":61,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0241,"confidence_interval":0.9409,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0168,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":667,"p95_latency_ms":815,"p99_latency_ms":874,"time_to_first_token_ms":148,"total_runtime_ms":741,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0258,"confidence_interval":0.9402,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9644,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":931,"time_to_first_token_ms":150,"total_runtime_ms":789,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0231,"confidence_interval":0.9373,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0863,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":742,"p95_latency_ms":907,"p99_latency_ms":973,"time_to_first_token_ms":143,"total_runtime_ms":824,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0201,"confidence_interval":0.9377,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3333,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":693,"p95_latency_ms":848,"p99_latency_ms":909,"time_to_first_token_ms":166,"total_runtime_ms":770,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0216,"confidence_interval":0.9484,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9329,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":743,"p95_latency_ms":908,"p99_latency_ms":974,"time_to_first_token_ms":163,"total_runtime_ms":826,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0241,"confidence_interval":0.9335,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3836,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":675,"p95_latency_ms":825,"p99_latency_ms":885,"time_to_first_token_ms":147,"total_runtime_ms":750,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0263,"confidence_interval":0.9489,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8424,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":679,"p95_latency_ms":830,"p99_latency_ms":891,"time_to_first_token_ms":169,"total_runtime_ms":755,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.014,"confidence_interval":0.9495,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8987,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":752,"p95_latency_ms":919,"p99_latency_ms":986,"time_to_first_token_ms":152,"total_runtime_ms":835,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0247,"confidence_interval":0.9362,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6844,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":714,"p95_latency_ms":873,"p99_latency_ms":937,"time_to_first_token_ms":149,"total_runtime_ms":794,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.936,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9121,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":731,"p95_latency_ms":893,"p99_latency_ms":958,"time_to_first_token_ms":144,"total_runtime_ms":812,"cold_start_penalty_ms":0,"peak_ram_mb":1157,"avg_ram_mb":771,"cpu_utilization_percent":52,"gpu_utilization_percent":61,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0143,"confidence_interval":0.9437,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2484,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":583,"p95_latency_ms":713,"p99_latency_ms":764,"time_to_first_token_ms":149,"total_runtime_ms":648,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0391,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0194,"confidence_interval":0.9405,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0137,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":519,"p95_latency_ms":634,"p99_latency_ms":680,"time_to_first_token_ms":128,"total_runtime_ms":577,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0156,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3082,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":590,"p95_latency_ms":721,"p99_latency_ms":774,"time_to_first_token_ms":136,"total_runtime_ms":656,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0487,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0105,"confidence_interval":0.9387,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2846,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":554,"p95_latency_ms":678,"p99_latency_ms":727,"time_to_first_token_ms":129,"total_runtime_ms":616,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.022,"confidence_interval":0.9348,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5553,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":524,"p95_latency_ms":640,"p99_latency_ms":687,"time_to_first_token_ms":143,"total_runtime_ms":582,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0112,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2394,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":530,"p95_latency_ms":648,"p99_latency_ms":695,"time_to_first_token_ms":141,"total_runtime_ms":589,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.022,"confidence_interval":0.9458,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.512,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":498,"p95_latency_ms":609,"p99_latency_ms":653,"time_to_first_token_ms":132,"total_runtime_ms":554,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0276,"confidence_interval":0.9332,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5719,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":541,"p95_latency_ms":662,"p99_latency_ms":710,"time_to_first_token_ms":138,"total_runtime_ms":602,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0342,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.027,"confidence_interval":0.9354,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.668,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":506,"p95_latency_ms":618,"p99_latency_ms":663,"time_to_first_token_ms":147,"total_runtime_ms":562,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0341,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0247,"confidence_interval":0.9421,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8128,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":578,"p95_latency_ms":707,"p99_latency_ms":758,"time_to_first_token_ms":127,"total_runtime_ms":643,"cold_start_penalty_ms":195,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0136,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.424,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":389,"p95_latency_ms":476,"p99_latency_ms":510,"time_to_first_token_ms":89,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0155,"confidence_interval":0.9484,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5882,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":424,"p95_latency_ms":518,"p99_latency_ms":556,"time_to_first_token_ms":99,"total_runtime_ms":471,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0309,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0111,"confidence_interval":0.943,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4955,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":362,"p95_latency_ms":442,"p99_latency_ms":474,"time_to_first_token_ms":103,"total_runtime_ms":402,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0173,"confidence_interval":0.9476,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5263,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":419,"p95_latency_ms":512,"p99_latency_ms":549,"time_to_first_token_ms":84,"total_runtime_ms":465,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0256,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7149,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":379,"p95_latency_ms":463,"p99_latency_ms":496,"time_to_first_token_ms":93,"total_runtime_ms":421,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0295,"confidence_interval":0.9395,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.2912,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":406,"p95_latency_ms":496,"p99_latency_ms":533,"time_to_first_token_ms":89,"total_runtime_ms":451,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0251,"confidence_interval":0.9384,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0554,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":365,"p95_latency_ms":446,"p99_latency_ms":478,"time_to_first_token_ms":79,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0214,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9554,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":377,"p95_latency_ms":461,"p99_latency_ms":494,"time_to_first_token_ms":77,"total_runtime_ms":419,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.05,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0153,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5033,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":338,"p95_latency_ms":413,"p99_latency_ms":443,"time_to_first_token_ms":98,"total_runtime_ms":375,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0299,"confidence_interval":0.9339,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3294,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":374,"p95_latency_ms":457,"p99_latency_ms":491,"time_to_first_token_ms":81,"total_runtime_ms":416,"cold_start_penalty_ms":0,"peak_ram_mb":567,"avg_ram_mb":378,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0173,"confidence_interval":0.9391,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.909,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":925,"p95_latency_ms":1131,"p99_latency_ms":1213,"time_to_first_token_ms":167,"total_runtime_ms":1028,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0286,"confidence_interval":0.9341,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0558,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":896,"p95_latency_ms":1095,"p99_latency_ms":1175,"time_to_first_token_ms":172,"total_runtime_ms":996,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0377,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0294,"confidence_interval":0.9328,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.452,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":929,"p95_latency_ms":1135,"p99_latency_ms":1218,"time_to_first_token_ms":188,"total_runtime_ms":1032,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.049,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0264,"confidence_interval":0.9429,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6304,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":878,"p95_latency_ms":1073,"p99_latency_ms":1151,"time_to_first_token_ms":184,"total_runtime_ms":975,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0148,"confidence_interval":0.9487,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.714,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":869,"p95_latency_ms":1062,"p99_latency_ms":1139,"time_to_first_token_ms":183,"total_runtime_ms":965,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0203,"confidence_interval":0.9333,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1242,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":839,"p95_latency_ms":1025,"p99_latency_ms":1100,"time_to_first_token_ms":195,"total_runtime_ms":932,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0118,"confidence_interval":0.9436,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1445,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":877,"p95_latency_ms":1072,"p99_latency_ms":1150,"time_to_first_token_ms":182,"total_runtime_ms":975,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.015,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3467,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":838,"p95_latency_ms":1024,"p99_latency_ms":1098,"time_to_first_token_ms":163,"total_runtime_ms":931,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0145,"confidence_interval":0.93,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.0552,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":846,"p95_latency_ms":1033,"p99_latency_ms":1109,"time_to_first_token_ms":171,"total_runtime_ms":939,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0223,"confidence_interval":0.9325,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2034,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.94,"schema_validity_rate":0.98,"parse_success_rate":0.93,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.06,"validation_failure_rate":0.03,"p50_latency_ms":839,"p95_latency_ms":1025,"p99_latency_ms":1100,"time_to_first_token_ms":191,"total_runtime_ms":932,"cold_start_penalty_ms":332,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":53,"gpu_utilization_percent":0,"oscillation_rate":0.0346,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0269,"confidence_interval":0.9446,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9943,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":611,"p95_latency_ms":747,"p99_latency_ms":801,"time_to_first_token_ms":142,"total_runtime_ms":679,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0447,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0156,"confidence_interval":0.9446,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5813,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":549,"p95_latency_ms":671,"p99_latency_ms":720,"time_to_first_token_ms":131,"total_runtime_ms":610,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0383,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.027,"confidence_interval":0.9371,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.9378,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":635,"p95_latency_ms":776,"p99_latency_ms":832,"time_to_first_token_ms":152,"total_runtime_ms":705,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0254,"confidence_interval":0.9357,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4996,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":541,"p95_latency_ms":661,"p99_latency_ms":709,"time_to_first_token_ms":143,"total_runtime_ms":601,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0228,"confidence_interval":0.9442,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1213,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":631,"p95_latency_ms":771,"p99_latency_ms":827,"time_to_first_token_ms":132,"total_runtime_ms":701,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0389,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0146,"confidence_interval":0.9321,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.054,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":625,"p95_latency_ms":764,"p99_latency_ms":819,"time_to_first_token_ms":125,"total_runtime_ms":694,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0379,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0278,"confidence_interval":0.9369,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7205,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":598,"p95_latency_ms":731,"p99_latency_ms":784,"time_to_first_token_ms":134,"total_runtime_ms":664,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0179,"confidence_interval":0.9499,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1911,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":755,"time_to_first_token_ms":120,"total_runtime_ms":640,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0147,"confidence_interval":0.9314,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2826,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":541,"p95_latency_ms":661,"p99_latency_ms":709,"time_to_first_token_ms":149,"total_runtime_ms":601,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.048,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0207,"confidence_interval":0.9413,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8693,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"long_context|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"long_context","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.97,"schema_validity_rate":0.98,"parse_success_rate":0.96,"retry_rate":0.05,"fallback_rate":0.01,"failure_rate":0.03,"validation_failure_rate":0.015,"p50_latency_ms":546,"p95_latency_ms":668,"p99_latency_ms":716,"time_to_first_token_ms":125,"total_runtime_ms":607,"cold_start_penalty_ms":0,"peak_ram_mb":964,"avg_ram_mb":643,"cpu_utilization_percent":43,"gpu_utilization_percent":0,"oscillation_rate":0.0481,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0294,"confidence_interval":0.939,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.7414,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":852,"p95_latency_ms":1041,"p99_latency_ms":1117,"time_to_first_token_ms":197,"total_runtime_ms":946,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0278,"confidence_interval":0.9382,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6503,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":827,"p95_latency_ms":1011,"p99_latency_ms":1084,"time_to_first_token_ms":195,"total_runtime_ms":919,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0267,"confidence_interval":0.9392,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6695,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":811,"p95_latency_ms":991,"p99_latency_ms":1064,"time_to_first_token_ms":192,"total_runtime_ms":901,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0381,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0286,"confidence_interval":0.941,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5675,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":799,"p95_latency_ms":977,"p99_latency_ms":1048,"time_to_first_token_ms":192,"total_runtime_ms":888,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0104,"confidence_interval":0.9311,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6173,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":750,"p95_latency_ms":917,"p99_latency_ms":983,"time_to_first_token_ms":213,"total_runtime_ms":833,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.021,"confidence_interval":0.9386,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6907,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":800,"p95_latency_ms":977,"p99_latency_ms":1048,"time_to_first_token_ms":199,"total_runtime_ms":888,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0462,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.014,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.6535,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":817,"p95_latency_ms":999,"p99_latency_ms":1071,"time_to_first_token_ms":217,"total_runtime_ms":908,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0272,"confidence_interval":0.945,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9487,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":798,"p95_latency_ms":975,"p99_latency_ms":1046,"time_to_first_token_ms":194,"total_runtime_ms":887,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0322,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0242,"confidence_interval":0.9456,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.7226,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":852,"p95_latency_ms":1042,"p99_latency_ms":1118,"time_to_first_token_ms":193,"total_runtime_ms":947,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0127,"confidence_interval":0.9457,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.1945,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":835,"p95_latency_ms":1020,"p99_latency_ms":1095,"time_to_first_token_ms":196,"total_runtime_ms":928,"cold_start_penalty_ms":416,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0359,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0211,"confidence_interval":0.9393,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.4027,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":446,"p95_latency_ms":545,"p99_latency_ms":585,"time_to_first_token_ms":175,"total_runtime_ms":496,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.025,"confidence_interval":0.9485,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.3759,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":389,"p95_latency_ms":476,"p99_latency_ms":510,"time_to_first_token_ms":172,"total_runtime_ms":432,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0123,"confidence_interval":0.9379,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.461,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":477,"p95_latency_ms":582,"p99_latency_ms":625,"time_to_first_token_ms":173,"total_runtime_ms":529,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0361,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0205,"confidence_interval":0.9439,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2311,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":406,"p95_latency_ms":496,"p99_latency_ms":532,"time_to_first_token_ms":154,"total_runtime_ms":451,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0255,"confidence_interval":0.9404,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2101,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":507,"time_to_first_token_ms":166,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0184,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5614,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":443,"p95_latency_ms":542,"p99_latency_ms":581,"time_to_first_token_ms":155,"total_runtime_ms":493,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0308,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0242,"confidence_interval":0.9371,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.0482,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":469,"p95_latency_ms":573,"p99_latency_ms":614,"time_to_first_token_ms":149,"total_runtime_ms":521,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.041,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.026,"confidence_interval":0.9313,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.5562,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":418,"p95_latency_ms":511,"p99_latency_ms":548,"time_to_first_token_ms":155,"total_runtime_ms":464,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0171,"confidence_interval":0.9406,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.335,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":411,"p95_latency_ms":502,"p99_latency_ms":539,"time_to_first_token_ms":174,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0318,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0233,"confidence_interval":0.934,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":6.2053,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":403,"p95_latency_ms":493,"p99_latency_ms":528,"time_to_first_token_ms":157,"total_runtime_ms":448,"cold_start_penalty_ms":0,"peak_ram_mb":896,"avg_ram_mb":672,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0476,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0175,"confidence_interval":0.9403,"validation_overhead_ms":32,"orchestration_overhead_ms":72,"event_dispatch_cost":5.9079,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1318,"p95_latency_ms":1610,"p99_latency_ms":1728,"time_to_first_token_ms":300,"total_runtime_ms":1464,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0214,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7625,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1304,"p95_latency_ms":1594,"p99_latency_ms":1710,"time_to_first_token_ms":300,"total_runtime_ms":1449,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0302,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0241,"confidence_interval":0.9307,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.147,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1334,"p95_latency_ms":1630,"p99_latency_ms":1749,"time_to_first_token_ms":308,"total_runtime_ms":1482,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0486,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.027,"confidence_interval":0.9493,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5197,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1338,"p95_latency_ms":1635,"p99_latency_ms":1754,"time_to_first_token_ms":306,"total_runtime_ms":1487,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0101,"confidence_interval":0.9499,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.7252,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1370,"p95_latency_ms":1675,"p99_latency_ms":1797,"time_to_first_token_ms":281,"total_runtime_ms":1523,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0223,"confidence_interval":0.9454,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8492,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1280,"p95_latency_ms":1564,"p99_latency_ms":1678,"time_to_first_token_ms":289,"total_runtime_ms":1422,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0496,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0147,"confidence_interval":0.9481,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9597,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1295,"p95_latency_ms":1583,"p99_latency_ms":1698,"time_to_first_token_ms":296,"total_runtime_ms":1439,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0286,"confidence_interval":0.9366,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9492,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1321,"p95_latency_ms":1615,"p99_latency_ms":1732,"time_to_first_token_ms":297,"total_runtime_ms":1468,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0457,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0226,"confidence_interval":0.9398,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3954,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1320,"p95_latency_ms":1613,"p99_latency_ms":1730,"time_to_first_token_ms":304,"total_runtime_ms":1466,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0353,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0207,"confidence_interval":0.9354,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.1364,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1303,"p95_latency_ms":1593,"p99_latency_ms":1708,"time_to_first_token_ms":299,"total_runtime_ms":1448,"cold_start_penalty_ms":707,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":61,"gpu_utilization_percent":0,"oscillation_rate":0.0436,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.92,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0261,"confidence_interval":0.9401,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5034,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":710,"p95_latency_ms":868,"p99_latency_ms":931,"time_to_first_token_ms":238,"total_runtime_ms":789,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0484,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0182,"confidence_interval":0.944,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.3545,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":714,"p95_latency_ms":873,"p99_latency_ms":937,"time_to_first_token_ms":241,"total_runtime_ms":794,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0278,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.898,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":660,"p95_latency_ms":807,"p99_latency_ms":866,"time_to_first_token_ms":244,"total_runtime_ms":734,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0384,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0153,"confidence_interval":0.9455,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.4239,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":674,"p95_latency_ms":823,"p99_latency_ms":883,"time_to_first_token_ms":265,"total_runtime_ms":748,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0292,"confidence_interval":0.933,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.8448,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":680,"p95_latency_ms":831,"p99_latency_ms":891,"time_to_first_token_ms":255,"total_runtime_ms":755,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0282,"confidence_interval":0.949,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":6.9721,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":709,"p95_latency_ms":867,"p99_latency_ms":930,"time_to_first_token_ms":243,"total_runtime_ms":788,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0472,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0164,"confidence_interval":0.9487,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.6875,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":744,"p95_latency_ms":910,"p99_latency_ms":976,"time_to_first_token_ms":234,"total_runtime_ms":827,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0229,"confidence_interval":0.9377,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.238,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":700,"p95_latency_ms":856,"p99_latency_ms":918,"time_to_first_token_ms":266,"total_runtime_ms":778,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0374,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0243,"confidence_interval":0.9402,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2232,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":700,"p95_latency_ms":856,"p99_latency_ms":918,"time_to_first_token_ms":249,"total_runtime_ms":778,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0273,"confidence_interval":0.9447,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.2206,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|low|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"low","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":642,"p95_latency_ms":784,"p99_latency_ms":841,"time_to_first_token_ms":257,"total_runtime_ms":713,"cold_start_penalty_ms":0,"peak_ram_mb":1523,"avg_ram_mb":1142,"cpu_utilization_percent":51,"gpu_utilization_percent":0,"oscillation_rate":0.044,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0127,"confidence_interval":0.9423,"validation_overhead_ms":38,"orchestration_overhead_ms":72,"event_dispatch_cost":7.5606,"memory_management_cost":9.84,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":590,"p95_latency_ms":721,"p99_latency_ms":774,"time_to_first_token_ms":170,"total_runtime_ms":656,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0172,"confidence_interval":0.9312,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.451,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":641,"p95_latency_ms":783,"p99_latency_ms":840,"time_to_first_token_ms":159,"total_runtime_ms":712,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0224,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.0365,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":660,"p95_latency_ms":806,"p99_latency_ms":865,"time_to_first_token_ms":158,"total_runtime_ms":733,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0489,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0143,"confidence_interval":0.9309,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.579,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":641,"p95_latency_ms":783,"p99_latency_ms":840,"time_to_first_token_ms":176,"total_runtime_ms":712,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.032,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0263,"confidence_interval":0.942,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2611,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":627,"p95_latency_ms":767,"p99_latency_ms":823,"time_to_first_token_ms":189,"total_runtime_ms":697,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0369,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0278,"confidence_interval":0.9469,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9056,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":576,"p95_latency_ms":704,"p99_latency_ms":756,"time_to_first_token_ms":172,"total_runtime_ms":640,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0257,"confidence_interval":0.9464,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.7551,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":664,"p95_latency_ms":812,"p99_latency_ms":871,"time_to_first_token_ms":186,"total_runtime_ms":738,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0448,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0292,"confidence_interval":0.9494,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6093,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":652,"p95_latency_ms":797,"p99_latency_ms":855,"time_to_first_token_ms":183,"total_runtime_ms":724,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0152,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9324,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":668,"p95_latency_ms":817,"p99_latency_ms":876,"time_to_first_token_ms":182,"total_runtime_ms":742,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0452,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0119,"confidence_interval":0.9488,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.978,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":665,"p95_latency_ms":812,"p99_latency_ms":872,"time_to_first_token_ms":162,"total_runtime_ms":739,"cold_start_penalty_ms":312,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0316,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0268,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3573,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":348,"p95_latency_ms":425,"p99_latency_ms":456,"time_to_first_token_ms":126,"total_runtime_ms":387,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0215,"confidence_interval":0.9491,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.4296,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":319,"p95_latency_ms":389,"p99_latency_ms":418,"time_to_first_token_ms":136,"total_runtime_ms":354,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0373,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0138,"confidence_interval":0.9342,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5193,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":290,"p95_latency_ms":354,"p99_latency_ms":380,"time_to_first_token_ms":133,"total_runtime_ms":322,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0488,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0297,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.6358,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":318,"p95_latency_ms":388,"p99_latency_ms":416,"time_to_first_token_ms":132,"total_runtime_ms":353,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0494,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0143,"confidence_interval":0.9466,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.9113,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":388,"p95_latency_ms":474,"p99_latency_ms":509,"time_to_first_token_ms":128,"total_runtime_ms":431,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0207,"confidence_interval":0.9487,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.5414,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":306,"p95_latency_ms":374,"p99_latency_ms":401,"time_to_first_token_ms":144,"total_runtime_ms":340,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0386,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0148,"confidence_interval":0.9341,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.692,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":507,"time_to_first_token_ms":144,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0303,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0219,"confidence_interval":0.9377,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":5.826,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":356,"p95_latency_ms":435,"p99_latency_ms":466,"time_to_first_token_ms":125,"total_runtime_ms":395,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0444,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.011,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3847,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":376,"p95_latency_ms":459,"p99_latency_ms":492,"time_to_first_token_ms":115,"total_runtime_ms":417,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0313,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0109,"confidence_interval":0.9365,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.3859,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":355,"p95_latency_ms":434,"p99_latency_ms":465,"time_to_first_token_ms":139,"total_runtime_ms":394,"cold_start_penalty_ms":0,"peak_ram_mb":672,"avg_ram_mb":504,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0282,"confidence_interval":0.9455,"validation_overhead_ms":32,"orchestration_overhead_ms":69,"event_dispatch_cost":6.2063,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":972,"p95_latency_ms":1188,"p99_latency_ms":1274,"time_to_first_token_ms":241,"total_runtime_ms":1080,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0266,"confidence_interval":0.938,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7913,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1057,"p95_latency_ms":1291,"p99_latency_ms":1385,"time_to_first_token_ms":241,"total_runtime_ms":1174,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0329,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0166,"confidence_interval":0.9388,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3191,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":996,"p95_latency_ms":1217,"p99_latency_ms":1306,"time_to_first_token_ms":251,"total_runtime_ms":1106,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0405,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0256,"confidence_interval":0.9339,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.5815,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1053,"p95_latency_ms":1287,"p99_latency_ms":1380,"time_to_first_token_ms":254,"total_runtime_ms":1170,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0358,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0206,"confidence_interval":0.9491,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.7563,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":964,"p95_latency_ms":1179,"p99_latency_ms":1264,"time_to_first_token_ms":243,"total_runtime_ms":1072,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0362,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0161,"confidence_interval":0.938,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.6786,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":967,"p95_latency_ms":1182,"p99_latency_ms":1268,"time_to_first_token_ms":245,"total_runtime_ms":1075,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0307,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.013,"confidence_interval":0.9426,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2325,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1052,"p95_latency_ms":1286,"p99_latency_ms":1380,"time_to_first_token_ms":228,"total_runtime_ms":1169,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0483,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0102,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.7052,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1036,"p95_latency_ms":1266,"p99_latency_ms":1359,"time_to_first_token_ms":242,"total_runtime_ms":1151,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0498,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0224,"confidence_interval":0.9363,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.0708,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1004,"p95_latency_ms":1228,"p99_latency_ms":1317,"time_to_first_token_ms":250,"total_runtime_ms":1116,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0137,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8201,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":1052,"p95_latency_ms":1285,"p99_latency_ms":1379,"time_to_first_token_ms":231,"total_runtime_ms":1168,"cold_start_penalty_ms":530,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":48,"gpu_utilization_percent":0,"oscillation_rate":0.0437,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0255,"confidence_interval":0.9431,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.407,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":548,"p95_latency_ms":670,"p99_latency_ms":718,"time_to_first_token_ms":186,"total_runtime_ms":609,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0459,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0116,"confidence_interval":0.9382,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1765,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":531,"p95_latency_ms":649,"p99_latency_ms":696,"time_to_first_token_ms":183,"total_runtime_ms":590,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0375,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0273,"confidence_interval":0.9395,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3092,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":537,"p95_latency_ms":657,"p99_latency_ms":705,"time_to_first_token_ms":206,"total_runtime_ms":597,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0317,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0174,"confidence_interval":0.9363,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.2482,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":478,"p95_latency_ms":584,"p99_latency_ms":627,"time_to_first_token_ms":208,"total_runtime_ms":531,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0372,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0112,"confidence_interval":0.933,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8085,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":551,"p95_latency_ms":674,"p99_latency_ms":723,"time_to_first_token_ms":182,"total_runtime_ms":612,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0247,"confidence_interval":0.9488,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.4904,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":550,"p95_latency_ms":672,"p99_latency_ms":721,"time_to_first_token_ms":213,"total_runtime_ms":611,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0468,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0232,"confidence_interval":0.9465,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":6.8468,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":516,"p95_latency_ms":630,"p99_latency_ms":676,"time_to_first_token_ms":196,"total_runtime_ms":573,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0315,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0242,"confidence_interval":0.9374,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1154,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":535,"p95_latency_ms":654,"p99_latency_ms":702,"time_to_first_token_ms":211,"total_runtime_ms":595,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0412,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0278,"confidence_interval":0.9418,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.7181,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":566,"p95_latency_ms":692,"p99_latency_ms":743,"time_to_first_token_ms":204,"total_runtime_ms":629,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0196,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.3252,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":579,"p95_latency_ms":708,"p99_latency_ms":759,"time_to_first_token_ms":206,"total_runtime_ms":643,"cold_start_penalty_ms":0,"peak_ram_mb":1142,"avg_ram_mb":857,"cpu_utilization_percent":38,"gpu_utilization_percent":0,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0181,"confidence_interval":0.9372,"validation_overhead_ms":38,"orchestration_overhead_ms":69,"event_dispatch_cost":7.1206,"memory_management_cost":8.88,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":455,"p95_latency_ms":557,"p99_latency_ms":597,"time_to_first_token_ms":158,"total_runtime_ms":506,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0253,"confidence_interval":0.9318,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8113,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":500,"p95_latency_ms":611,"p99_latency_ms":655,"time_to_first_token_ms":155,"total_runtime_ms":555,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0344,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0288,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2343,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":481,"p95_latency_ms":588,"p99_latency_ms":631,"time_to_first_token_ms":151,"total_runtime_ms":535,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0339,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0249,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8118,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":476,"p95_latency_ms":582,"p99_latency_ms":624,"time_to_first_token_ms":134,"total_runtime_ms":529,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0414,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.011,"confidence_interval":0.9462,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0791,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":426,"p95_latency_ms":521,"p99_latency_ms":559,"time_to_first_token_ms":138,"total_runtime_ms":473,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0245,"confidence_interval":0.9425,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.2707,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":512,"p95_latency_ms":626,"p99_latency_ms":671,"time_to_first_token_ms":145,"total_runtime_ms":569,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0328,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0284,"confidence_interval":0.9316,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9589,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":459,"p95_latency_ms":561,"p99_latency_ms":602,"time_to_first_token_ms":165,"total_runtime_ms":510,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0434,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0102,"confidence_interval":0.9334,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0039,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":456,"p95_latency_ms":558,"p99_latency_ms":598,"time_to_first_token_ms":156,"total_runtime_ms":507,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0215,"confidence_interval":0.9411,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6658,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":425,"p95_latency_ms":519,"p99_latency_ms":557,"time_to_first_token_ms":146,"total_runtime_ms":472,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0198,"confidence_interval":0.9328,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4501,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":435,"p95_latency_ms":532,"p99_latency_ms":571,"time_to_first_token_ms":149,"total_runtime_ms":484,"cold_start_penalty_ms":234,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":39,"gpu_utilization_percent":20,"oscillation_rate":0.031,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0225,"confidence_interval":0.9499,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3679,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":219,"p95_latency_ms":268,"p99_latency_ms":287,"time_to_first_token_ms":117,"total_runtime_ms":244,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.033,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0174,"confidence_interval":0.9304,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.6249,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":212,"p95_latency_ms":260,"p99_latency_ms":278,"time_to_first_token_ms":90,"total_runtime_ms":236,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0324,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0213,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.5865,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":221,"p95_latency_ms":270,"p99_latency_ms":290,"time_to_first_token_ms":111,"total_runtime_ms":245,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.043,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0109,"confidence_interval":0.9426,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.7013,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":310,"p95_latency_ms":379,"p99_latency_ms":407,"time_to_first_token_ms":101,"total_runtime_ms":345,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0409,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0132,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3001,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":217,"p95_latency_ms":265,"p99_latency_ms":284,"time_to_first_token_ms":109,"total_runtime_ms":241,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.037,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0191,"confidence_interval":0.9449,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.4012,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":273,"p95_latency_ms":334,"p99_latency_ms":358,"time_to_first_token_ms":107,"total_runtime_ms":304,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0428,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0119,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.0941,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":264,"p95_latency_ms":322,"p99_latency_ms":346,"time_to_first_token_ms":108,"total_runtime_ms":293,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0417,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0224,"confidence_interval":0.9319,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.8988,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":214,"p95_latency_ms":262,"p99_latency_ms":281,"time_to_first_token_ms":94,"total_runtime_ms":238,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0432,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0159,"confidence_interval":0.9355,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.3448,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":232,"p95_latency_ms":284,"p99_latency_ms":305,"time_to_first_token_ms":115,"total_runtime_ms":258,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0138,"confidence_interval":0.9415,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":5.9461,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":313,"p95_latency_ms":383,"p99_latency_ms":411,"time_to_first_token_ms":119,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":504,"avg_ram_mb":378,"cpu_utilization_percent":29,"gpu_utilization_percent":20,"oscillation_rate":0.0449,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0136,"confidence_interval":0.9463,"validation_overhead_ms":32,"orchestration_overhead_ms":66,"event_dispatch_cost":6.277,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":741,"p95_latency_ms":905,"p99_latency_ms":971,"time_to_first_token_ms":209,"total_runtime_ms":823,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0239,"confidence_interval":0.9334,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2933,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":766,"p95_latency_ms":936,"p99_latency_ms":1004,"time_to_first_token_ms":183,"total_runtime_ms":851,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.01,"confidence_interval":0.9487,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.3348,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":762,"p95_latency_ms":931,"p99_latency_ms":999,"time_to_first_token_ms":210,"total_runtime_ms":847,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0273,"confidence_interval":0.9439,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2385,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":807,"p95_latency_ms":987,"p99_latency_ms":1059,"time_to_first_token_ms":206,"total_runtime_ms":897,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0493,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0264,"confidence_interval":0.9398,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6122,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":813,"p95_latency_ms":994,"p99_latency_ms":1066,"time_to_first_token_ms":199,"total_runtime_ms":904,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0202,"confidence_interval":0.9305,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6568,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":807,"p95_latency_ms":986,"p99_latency_ms":1058,"time_to_first_token_ms":185,"total_runtime_ms":896,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0304,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0182,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.1865,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":804,"p95_latency_ms":983,"p99_latency_ms":1055,"time_to_first_token_ms":199,"total_runtime_ms":894,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.035,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0222,"confidence_interval":0.9309,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.9071,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":739,"p95_latency_ms":903,"p99_latency_ms":969,"time_to_first_token_ms":208,"total_runtime_ms":821,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.046,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0253,"confidence_interval":0.9355,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4247,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":719,"p95_latency_ms":879,"p99_latency_ms":943,"time_to_first_token_ms":205,"total_runtime_ms":799,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0222,"confidence_interval":0.9373,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0337,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":821,"p95_latency_ms":1003,"p99_latency_ms":1076,"time_to_first_token_ms":200,"total_runtime_ms":912,"cold_start_penalty_ms":398,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":39,"gpu_utilization_percent":34,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0104,"confidence_interval":0.9349,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.6506,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":380,"p95_latency_ms":464,"p99_latency_ms":498,"time_to_first_token_ms":151,"total_runtime_ms":422,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0101,"confidence_interval":0.9318,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4733,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":388,"p95_latency_ms":474,"p99_latency_ms":508,"time_to_first_token_ms":141,"total_runtime_ms":431,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0419,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0255,"confidence_interval":0.9446,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0293,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":423,"p95_latency_ms":518,"p99_latency_ms":555,"time_to_first_token_ms":171,"total_runtime_ms":470,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0435,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0292,"confidence_interval":0.9366,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8402,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":384,"p95_latency_ms":469,"p99_latency_ms":503,"time_to_first_token_ms":162,"total_runtime_ms":426,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0193,"confidence_interval":0.9371,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.4907,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":435,"p95_latency_ms":531,"p99_latency_ms":570,"time_to_first_token_ms":142,"total_runtime_ms":483,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0306,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0258,"confidence_interval":0.9364,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.7663,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":508,"time_to_first_token_ms":151,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0295,"confidence_interval":0.9333,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.456,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":387,"p95_latency_ms":473,"p99_latency_ms":507,"time_to_first_token_ms":138,"total_runtime_ms":430,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0335,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0184,"confidence_interval":0.9456,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.2753,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":410,"p95_latency_ms":501,"p99_latency_ms":538,"time_to_first_token_ms":138,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0478,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0133,"confidence_interval":0.9398,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":6.8025,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":411,"p95_latency_ms":502,"p99_latency_ms":539,"time_to_first_token_ms":167,"total_runtime_ms":456,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0323,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0164,"confidence_interval":0.9319,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.0096,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"medium_gpu","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":374,"p95_latency_ms":457,"p99_latency_ms":491,"time_to_first_token_ms":162,"total_runtime_ms":416,"cold_start_penalty_ms":0,"peak_ram_mb":857,"avg_ram_mb":643,"cpu_utilization_percent":29,"gpu_utilization_percent":34,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.025,"confidence_interval":0.9477,"validation_overhead_ms":38,"orchestration_overhead_ms":66,"event_dispatch_cost":7.61,"memory_management_cost":8.16,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":401,"p95_latency_ms":490,"p99_latency_ms":526,"time_to_first_token_ms":123,"total_runtime_ms":446,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0413,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0183,"confidence_interval":0.939,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.92,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":396,"p95_latency_ms":484,"p99_latency_ms":519,"time_to_first_token_ms":128,"total_runtime_ms":440,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0169,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7818,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":405,"p95_latency_ms":495,"p99_latency_ms":531,"time_to_first_token_ms":147,"total_runtime_ms":450,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0363,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0159,"confidence_interval":0.9447,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8335,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":386,"p95_latency_ms":472,"p99_latency_ms":507,"time_to_first_token_ms":153,"total_runtime_ms":429,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0147,"confidence_interval":0.9383,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.575,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":444,"p95_latency_ms":542,"p99_latency_ms":582,"time_to_first_token_ms":124,"total_runtime_ms":493,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0397,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0247,"confidence_interval":0.9442,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4067,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":448,"p95_latency_ms":548,"p99_latency_ms":588,"time_to_first_token_ms":134,"total_runtime_ms":498,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0464,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0229,"confidence_interval":0.9437,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.7982,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":381,"p95_latency_ms":466,"p99_latency_ms":500,"time_to_first_token_ms":145,"total_runtime_ms":424,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0418,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.024,"confidence_interval":0.932,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.719,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":388,"p95_latency_ms":474,"p99_latency_ms":508,"time_to_first_token_ms":134,"total_runtime_ms":431,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0271,"confidence_interval":0.9435,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.8923,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":453,"p95_latency_ms":553,"p99_latency_ms":593,"time_to_first_token_ms":137,"total_runtime_ms":503,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0102,"confidence_interval":0.9436,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.5602,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":456,"p95_latency_ms":557,"p99_latency_ms":598,"time_to_first_token_ms":130,"total_runtime_ms":507,"cold_start_penalty_ms":195,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0396,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0185,"confidence_interval":0.9492,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4676,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":219,"p95_latency_ms":268,"p99_latency_ms":288,"time_to_first_token_ms":92,"total_runtime_ms":244,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0382,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0228,"confidence_interval":0.9408,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0873,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":214,"p95_latency_ms":261,"p99_latency_ms":280,"time_to_first_token_ms":110,"total_runtime_ms":237,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0416,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0173,"confidence_interval":0.9335,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6281,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":196,"p95_latency_ms":240,"p99_latency_ms":257,"time_to_first_token_ms":106,"total_runtime_ms":218,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.04,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0186,"confidence_interval":0.9444,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4363,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":266,"p95_latency_ms":326,"p99_latency_ms":349,"time_to_first_token_ms":105,"total_runtime_ms":296,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0422,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0291,"confidence_interval":0.9363,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.6803,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":209,"p95_latency_ms":255,"p99_latency_ms":274,"time_to_first_token_ms":84,"total_runtime_ms":232,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0326,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0293,"confidence_interval":0.9333,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0441,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":177,"p95_latency_ms":217,"p99_latency_ms":232,"time_to_first_token_ms":86,"total_runtime_ms":197,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0312,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.014,"confidence_interval":0.9475,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3725,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":223,"p95_latency_ms":272,"p99_latency_ms":292,"time_to_first_token_ms":81,"total_runtime_ms":247,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0446,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0282,"confidence_interval":0.9497,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.3379,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":205,"p95_latency_ms":250,"p99_latency_ms":268,"time_to_first_token_ms":106,"total_runtime_ms":228,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0398,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0202,"confidence_interval":0.931,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.0549,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":226,"p95_latency_ms":276,"p99_latency_ms":296,"time_to_first_token_ms":92,"total_runtime_ms":251,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0461,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.015,"confidence_interval":0.947,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":6.4594,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e2b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e2b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":251,"p95_latency_ms":307,"p99_latency_ms":329,"time_to_first_token_ms":92,"total_runtime_ms":279,"cold_start_penalty_ms":0,"peak_ram_mb":420,"avg_ram_mb":315,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0492,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0236,"confidence_interval":0.9401,"validation_overhead_ms":32,"orchestration_overhead_ms":65,"event_dispatch_cost":5.9269,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":0,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":638,"p95_latency_ms":779,"p99_latency_ms":836,"time_to_first_token_ms":194,"total_runtime_ms":708,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0366,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0282,"confidence_interval":0.9304,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.2386,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":1,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":649,"p95_latency_ms":793,"p99_latency_ms":851,"time_to_first_token_ms":180,"total_runtime_ms":721,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0371,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0251,"confidence_interval":0.9397,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8554,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":2,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":671,"p95_latency_ms":820,"p99_latency_ms":879,"time_to_first_token_ms":192,"total_runtime_ms":745,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0365,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0135,"confidence_interval":0.9444,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5467,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":3,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":697,"p95_latency_ms":852,"p99_latency_ms":914,"time_to_first_token_ms":165,"total_runtime_ms":774,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0133,"confidence_interval":0.9419,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8604,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":4,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":674,"p95_latency_ms":824,"p99_latency_ms":883,"time_to_first_token_ms":185,"total_runtime_ms":749,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0399,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0252,"confidence_interval":0.9313,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.379,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":5,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":648,"p95_latency_ms":791,"p99_latency_ms":849,"time_to_first_token_ms":165,"total_runtime_ms":719,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0477,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0146,"confidence_interval":0.942,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3985,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":6,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":640,"p95_latency_ms":782,"p99_latency_ms":839,"time_to_first_token_ms":168,"total_runtime_ms":711,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0403,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0114,"confidence_interval":0.9373,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8167,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":7,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":604,"p95_latency_ms":739,"p99_latency_ms":792,"time_to_first_token_ms":165,"total_runtime_ms":672,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0425,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0252,"confidence_interval":0.935,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1087,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":8,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":687,"p95_latency_ms":840,"p99_latency_ms":901,"time_to_first_token_ms":171,"total_runtime_ms":764,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.0471,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0212,"confidence_interval":0.9332,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5305,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|cold","run":9,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"cold","task_success_rate":0.88,"schema_validity_rate":0.98,"parse_success_rate":0.86,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.12,"validation_failure_rate":0.06,"p50_latency_ms":617,"p95_latency_ms":755,"p99_latency_ms":809,"time_to_first_token_ms":172,"total_runtime_ms":686,"cold_start_penalty_ms":332,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":34,"gpu_utilization_percent":0,"oscillation_rate":0.046,"quarantine_trigger_rate":0.12,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0212,"confidence_interval":0.9301,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6336,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":0,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":301,"p95_latency_ms":368,"p99_latency_ms":395,"time_to_first_token_ms":136,"total_runtime_ms":335,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0463,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":1,"variance_across_runs":0.0182,"confidence_interval":0.9414,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.4983,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":1,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":396,"p95_latency_ms":484,"p99_latency_ms":519,"time_to_first_token_ms":139,"total_runtime_ms":440,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0438,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9988,"variance_across_runs":0.0225,"confidence_interval":0.9369,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6418,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":2,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":330,"p95_latency_ms":403,"p99_latency_ms":432,"time_to_first_token_ms":133,"total_runtime_ms":366,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0458,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9975,"variance_across_runs":0.0127,"confidence_interval":0.9446,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5982,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":3,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":318,"p95_latency_ms":389,"p99_latency_ms":417,"time_to_first_token_ms":141,"total_runtime_ms":354,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0388,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9962,"variance_across_runs":0.0146,"confidence_interval":0.931,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8179,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":4,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":328,"p95_latency_ms":401,"p99_latency_ms":430,"time_to_first_token_ms":135,"total_runtime_ms":364,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0336,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.995,"variance_across_runs":0.0254,"confidence_interval":0.9496,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":6.8461,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":5,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":364,"p95_latency_ms":445,"p99_latency_ms":478,"time_to_first_token_ms":142,"total_runtime_ms":405,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0433,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9938,"variance_across_runs":0.0198,"confidence_interval":0.9323,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.1129,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":6,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":331,"p95_latency_ms":405,"p99_latency_ms":434,"time_to_first_token_ms":151,"total_runtime_ms":368,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0454,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9925,"variance_across_runs":0.0293,"confidence_interval":0.9353,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.6864,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":7,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":314,"p95_latency_ms":383,"p99_latency_ms":411,"time_to_first_token_ms":123,"total_runtime_ms":348,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0443,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9912,"variance_across_runs":0.0162,"confidence_interval":0.941,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5063,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":8,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":357,"p95_latency_ms":437,"p99_latency_ms":469,"time_to_first_token_ms":143,"total_runtime_ms":397,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0406,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.99,"variance_across_runs":0.0136,"confidence_interval":0.937,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.3265,"memory_management_cost":7.8,"memory_pressure_events":0} +{"combo_id":"noise_input|high|gemma4:e4b-it-qat|system_selected|warm","run":9,"scenario":"noise_input","hardware_profile":"high","model":"gemma4:e4b-it-qat","selection_mode":"system_selected","cold_or_warm":"warm","task_success_rate":0.91,"schema_validity_rate":0.98,"parse_success_rate":0.89,"retry_rate":0.1,"fallback_rate":0.01,"failure_rate":0.09,"validation_failure_rate":0.045,"p50_latency_ms":312,"p95_latency_ms":382,"p99_latency_ms":410,"time_to_first_token_ms":118,"total_runtime_ms":347,"cold_start_penalty_ms":0,"peak_ram_mb":714,"avg_ram_mb":536,"cpu_utilization_percent":24,"gpu_utilization_percent":0,"oscillation_rate":0.0355,"quarantine_trigger_rate":0.02,"smallest_sufficient_hit_rate":0.35,"overprovision_rate":0.04,"underprovision_rate":0.03,"overlay_activation_accuracy":1,"latency_hint_accuracy":0.88,"unexplained_wait_rate":0.06,"determinism_score":0.9888,"variance_across_runs":0.0106,"confidence_interval":0.9492,"validation_overhead_ms":38,"orchestration_overhead_ms":65,"event_dispatch_cost":7.5479,"memory_management_cost":7.8,"memory_pressure_events":0} diff --git a/reports/repo_plan.json b/reports/repo_plan.json index 71d07da..52dfaa6 100644 --- a/reports/repo_plan.json +++ b/reports/repo_plan.json @@ -1,5 +1,5 @@ { - "generated_at": "2026-08-18T19:23:33.758Z", + "generated_at": "2026-08-25T17:30:23.118Z", "input_document": "selectpilot_monolith_v3.json", "phases": [ { diff --git a/reports/runtime_policy_validation.json b/reports/runtime_policy_validation.json index 44e54e4..6bbe474 100644 --- a/reports/runtime_policy_validation.json +++ b/reports/runtime_policy_validation.json @@ -15,8 +15,8 @@ "fallback_cycle_detected", "quarantined_model_promoted" ], - "policy_version": "runtime-policy-c44de2e4a8a1", - "checked_at_unix_ms": 1700798162081, + "policy_version": "runtime-policy-8226a11e6ae3", + "checked_at_unix_ms": 1700423493347, "policy_path": "runtime/model_policy.json", "runtime_registry_path": "runtime/model_registry.runtime.json" } diff --git a/reports/runtime_selection_summary.json b/reports/runtime_selection_summary.json index 2f0755c..ee31a45 100644 --- a/reports/runtime_selection_summary.json +++ b/reports/runtime_selection_summary.json @@ -1,9 +1,10 @@ { - "generated_at_unix_ms": 1700798162081, - "policy_version": "runtime-policy-c44de2e4a8a1", + "generated_at_unix_ms": 1700423493347, + "policy_version": "runtime-policy-8226a11e6ae3", "defaults_count": 12, "preferred_models": [ - "qwen2.5:0.5b" + "gemma4:e2b-it-qat", + "gemma4:e4b-it-qat" ], "quarantined_models": [], "rejected_models": [], diff --git a/reports/scenario_results.json b/reports/scenario_results.json index 1c4c2c3..f41fe48 100644 --- a/reports/scenario_results.json +++ b/reports/scenario_results.json @@ -3,10 +3,10 @@ "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:0.5b|system_selected|cold", + "combo_id": "rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -14,19 +14,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 994, - "total_runtime_ms": 903, - "weighted_score": 0.8378 + "p95_latency_ms": 976, + "total_runtime_ms": 887, + "weighted_score": 0.8395 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:0.5b|system_selected|warm", + "combo_id": "rewrite_minimal|low|gemma4:e2b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -34,19 +34,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 533, - "total_runtime_ms": 485, - "weighted_score": 0.9018 + "p95_latency_ms": 519, + "total_runtime_ms": 471, + "weighted_score": 0.9033 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:1.5b|system_selected|cold", + "combo_id": "rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -54,19 +54,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1303, - "total_runtime_ms": 1184, - "weighted_score": 0.7974 + "p95_latency_ms": 1615, + "total_runtime_ms": 1468, + "weighted_score": 0.7569 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:1.5b|system_selected|warm", + "combo_id": "rewrite_minimal|low|gemma4:e4b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -74,19 +74,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 679, - "total_runtime_ms": 618, - "weighted_score": 0.8778 + "p95_latency_ms": 819, + "total_runtime_ms": 744, + "weighted_score": 0.8545 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:7b|system_selected|cold", + "combo_id": "rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:7b", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -94,19 +94,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 2519, - "total_runtime_ms": 2290, - "weighted_score": 0.6414 + "p95_latency_ms": 750, + "total_runtime_ms": 682, + "weighted_score": 0.8689 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:7b|system_selected|warm", + "combo_id": "rewrite_minimal|medium|gemma4:e2b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:7b", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -114,59 +114,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1312, - "total_runtime_ms": 1193, - "weighted_score": 0.7783 + "p95_latency_ms": 387, + "total_runtime_ms": 351, + "weighted_score": 0.9232 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:14b|system_selected|cold", + "combo_id": "rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:14b", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3732, - "total_runtime_ms": 3392, - "weighted_score": 0.5885 + "failure_rate": 0.06, + "p95_latency_ms": 1218, + "total_runtime_ms": 1108, + "weighted_score": 0.8079 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|low|qwen2:14b|system_selected|warm", + "combo_id": "rewrite_minimal|medium|gemma4:e4b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", - "hardware_profile": "low", - "model": "qwen2:14b", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1880, - "total_runtime_ms": 1709, - "weighted_score": 0.6685 + "failure_rate": 0.03, + "p95_latency_ms": 640, + "total_runtime_ms": 582, + "weighted_score": 0.8838 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:0.5b|system_selected|cold", + "combo_id": "rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -174,19 +174,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 756, - "total_runtime_ms": 687, - "weighted_score": 0.8682 + "p95_latency_ms": 572, + "total_runtime_ms": 520, + "weighted_score": 0.8917 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:0.5b|system_selected|warm", + "combo_id": "rewrite_minimal|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -194,19 +194,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 416, - "total_runtime_ms": 378, - "weighted_score": 0.9202 + "p95_latency_ms": 324, + "total_runtime_ms": 294, + "weighted_score": 0.9345 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:1.5b|system_selected|cold", + "combo_id": "rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -214,19 +214,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 994, - "total_runtime_ms": 904, - "weighted_score": 0.8373 + "p95_latency_ms": 936, + "total_runtime_ms": 850, + "weighted_score": 0.8447 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:1.5b|system_selected|warm", + "combo_id": "rewrite_minimal|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -234,19 +234,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 518, - "total_runtime_ms": 471, - "weighted_score": 0.903 + "p95_latency_ms": 514, + "total_runtime_ms": 467, + "weighted_score": 0.9049 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:7b|system_selected|cold", + "combo_id": "rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:7b", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -254,19 +254,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1921, - "total_runtime_ms": 1746, - "weighted_score": 0.7175 + "p95_latency_ms": 501, + "total_runtime_ms": 456, + "weighted_score": 0.9013 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:7b|system_selected|warm", + "combo_id": "rewrite_minimal|high|gemma4:e2b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:7b", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -274,59 +274,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 989, - "total_runtime_ms": 899, - "weighted_score": 0.8287 + "p95_latency_ms": 282, + "total_runtime_ms": 256, + "weighted_score": 0.9412 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:14b|system_selected|cold", + "combo_id": "rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|cold", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:14b", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2789, - "total_runtime_ms": 2536, - "weighted_score": 0.6154 + "failure_rate": 0.06, + "p95_latency_ms": 798, + "total_runtime_ms": 725, + "weighted_score": 0.8628 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium|qwen2:14b|system_selected|warm", + "combo_id": "rewrite_minimal|high|gemma4:e4b-it-qat|system_selected|warm", "scenario": "rewrite_minimal", - "hardware_profile": "medium", - "model": "qwen2:14b", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1437, - "total_runtime_ms": 1306, - "weighted_score": 0.7396 + "failure_rate": 0.03, + "p95_latency_ms": 418, + "total_runtime_ms": 380, + "weighted_score": 0.9188 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -334,19 +334,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 567, - "total_runtime_ms": 515, - "weighted_score": 0.8922 + "p95_latency_ms": 1078, + "total_runtime_ms": 980, + "weighted_score": 0.8293 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "rewrite_constrained|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -354,19 +354,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 319, - "total_runtime_ms": 290, - "weighted_score": 0.935 + "p95_latency_ms": 600, + "total_runtime_ms": 546, + "weighted_score": 0.8951 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -374,19 +374,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 765, - "total_runtime_ms": 695, - "weighted_score": 0.8671 + "p95_latency_ms": 1788, + "total_runtime_ms": 1625, + "weighted_score": 0.7396 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "rewrite_constrained|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -394,19 +394,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 419, - "total_runtime_ms": 381, - "weighted_score": 0.9197 + "p95_latency_ms": 996, + "total_runtime_ms": 905, + "weighted_score": 0.8367 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -414,19 +414,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1461, - "total_runtime_ms": 1328, - "weighted_score": 0.7771 + "p95_latency_ms": 813, + "total_runtime_ms": 740, + "weighted_score": 0.8625 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "rewrite_constrained|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -434,59 +434,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 783, - "total_runtime_ms": 712, - "weighted_score": 0.8628 + "p95_latency_ms": 473, + "total_runtime_ms": 430, + "weighted_score": 0.9145 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2130, - "total_runtime_ms": 1936, - "weighted_score": 0.6725 + "failure_rate": 0.06, + "p95_latency_ms": 1334, + "total_runtime_ms": 1213, + "weighted_score": 0.7964 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "rewrite_minimal", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "rewrite_constrained|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1093, - "total_runtime_ms": 994, - "weighted_score": 0.7942 + "failure_rate": 0.03, + "p95_latency_ms": 780, + "total_runtime_ms": 709, + "weighted_score": 0.8697 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:0.5b|system_selected|cold", - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -494,19 +494,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 494, - "total_runtime_ms": 449, - "weighted_score": 0.902 + "p95_latency_ms": 650, + "total_runtime_ms": 591, + "weighted_score": 0.8839 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:0.5b|system_selected|warm", - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "rewrite_constrained|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -514,19 +514,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 277, - "total_runtime_ms": 252, - "weighted_score": 0.9417 + "p95_latency_ms": 379, + "total_runtime_ms": 344, + "weighted_score": 0.929 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:1.5b|system_selected|cold", - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -534,19 +534,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 663, - "total_runtime_ms": 602, - "weighted_score": 0.8807 + "p95_latency_ms": 1031, + "total_runtime_ms": 937, + "weighted_score": 0.8352 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:1.5b|system_selected|warm", - "scenario": "rewrite_minimal", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "rewrite_constrained|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -554,19 +554,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 346, - "total_runtime_ms": 314, - "weighted_score": 0.9304 + "p95_latency_ms": 584, + "total_runtime_ms": 531, + "weighted_score": 0.8979 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:7b|system_selected|cold", - "scenario": "rewrite_minimal", + "combo_id": "rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -574,19 +574,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1235, - "total_runtime_ms": 1123, - "weighted_score": 0.8065 + "p95_latency_ms": 547, + "total_runtime_ms": 497, + "weighted_score": 0.8967 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:7b|system_selected|warm", - "scenario": "rewrite_minimal", + "combo_id": "rewrite_constrained|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -594,59 +594,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 662, - "total_runtime_ms": 602, - "weighted_score": 0.8818 + "p95_latency_ms": 314, + "total_runtime_ms": 286, + "weighted_score": 0.938 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:14b|system_selected|cold", - "scenario": "rewrite_minimal", + "combo_id": "rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "rewrite_constrained", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1798, - "total_runtime_ms": 1635, - "weighted_score": 0.7158 + "failure_rate": 0.06, + "p95_latency_ms": 878, + "total_runtime_ms": 798, + "weighted_score": 0.8548 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_minimal|high|qwen2:14b|system_selected|warm", - "scenario": "rewrite_minimal", + "combo_id": "rewrite_constrained|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "rewrite_constrained", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 932, - "total_runtime_ms": 847, - "weighted_score": 0.8204 + "failure_rate": 0.03, + "p95_latency_ms": 495, + "total_runtime_ms": 450, + "weighted_score": 0.9111 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:0.5b|system_selected|cold", - "scenario": "rewrite_constrained", + "combo_id": "classification_binary|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_binary", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -654,19 +654,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1091, - "total_runtime_ms": 991, - "weighted_score": 0.8281 + "p95_latency_ms": 970, + "total_runtime_ms": 882, + "weighted_score": 0.8401 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:0.5b|system_selected|warm", - "scenario": "rewrite_constrained", + "combo_id": "classification_binary|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_binary", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -674,19 +674,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 611, - "total_runtime_ms": 555, - "weighted_score": 0.8941 + "p95_latency_ms": 537, + "total_runtime_ms": 489, + "weighted_score": 0.9014 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:1.5b|system_selected|cold", - "scenario": "rewrite_constrained", + "combo_id": "classification_binary|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_binary", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -694,19 +694,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1431, - "total_runtime_ms": 1301, - "weighted_score": 0.7846 + "p95_latency_ms": 1621, + "total_runtime_ms": 1474, + "weighted_score": 0.7562 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:1.5b|system_selected|warm", - "scenario": "rewrite_constrained", + "combo_id": "classification_binary|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_binary", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -714,19 +714,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 798, - "total_runtime_ms": 726, - "weighted_score": 0.8659 + "p95_latency_ms": 845, + "total_runtime_ms": 768, + "weighted_score": 0.8518 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:7b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "classification_binary|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_binary", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -734,19 +734,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 2802, - "total_runtime_ms": 2547, - "weighted_score": 0.6414 + "p95_latency_ms": 748, + "total_runtime_ms": 680, + "weighted_score": 0.869 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:7b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "classification_binary|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_binary", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -754,59 +754,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1535, - "total_runtime_ms": 1396, - "weighted_score": 0.7559 + "p95_latency_ms": 392, + "total_runtime_ms": 356, + "weighted_score": 0.9227 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:14b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "classification_binary|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_binary", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 4087, - "total_runtime_ms": 3715, - "weighted_score": 0.5885 + "failure_rate": 0.06, + "p95_latency_ms": 1213, + "total_runtime_ms": 1103, + "weighted_score": 0.8084 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|low|qwen2:14b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "classification_binary|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_binary", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2253, - "total_runtime_ms": 2048, - "weighted_score": 0.6312 + "failure_rate": 0.03, + "p95_latency_ms": 637, + "total_runtime_ms": 579, + "weighted_score": 0.884 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:0.5b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_binary", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -814,19 +814,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 845, - "total_runtime_ms": 768, - "weighted_score": 0.8594 + "p95_latency_ms": 620, + "total_runtime_ms": 563, + "weighted_score": 0.8869 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:0.5b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "classification_binary|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_binary", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -834,19 +834,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 478, - "total_runtime_ms": 434, - "weighted_score": 0.9141 + "p95_latency_ms": 334, + "total_runtime_ms": 303, + "weighted_score": 0.9335 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:1.5b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_binary", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -854,19 +854,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1093, - "total_runtime_ms": 994, - "weighted_score": 0.8275 + "p95_latency_ms": 935, + "total_runtime_ms": 850, + "weighted_score": 0.8447 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:1.5b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "classification_binary|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_binary", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -874,19 +874,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 619, - "total_runtime_ms": 563, - "weighted_score": 0.8929 + "p95_latency_ms": 495, + "total_runtime_ms": 450, + "weighted_score": 0.9068 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:7b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "classification_binary|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -894,19 +894,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 2126, - "total_runtime_ms": 1933, - "weighted_score": 0.697 + "p95_latency_ms": 498, + "total_runtime_ms": 452, + "weighted_score": 0.9016 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:7b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "classification_binary|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -914,59 +914,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1167, - "total_runtime_ms": 1061, - "weighted_score": 0.8109 + "p95_latency_ms": 302, + "total_runtime_ms": 275, + "weighted_score": 0.9392 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:14b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "classification_binary|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3087, - "total_runtime_ms": 2806, - "weighted_score": 0.6154 + "failure_rate": 0.06, + "p95_latency_ms": 802, + "total_runtime_ms": 729, + "weighted_score": 0.8624 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium|qwen2:14b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "classification_binary|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_binary", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1715, - "total_runtime_ms": 1559, - "weighted_score": 0.7119 + "failure_rate": 0.03, + "p95_latency_ms": 446, + "total_runtime_ms": 406, + "weighted_score": 0.9159 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "classification_multiclass|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -974,19 +974,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 635, - "total_runtime_ms": 578, - "weighted_score": 0.8853 + "p95_latency_ms": 1074, + "total_runtime_ms": 977, + "weighted_score": 0.8297 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "classification_multiclass|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -994,19 +994,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 387, - "total_runtime_ms": 352, - "weighted_score": 0.9282 + "p95_latency_ms": 604, + "total_runtime_ms": 549, + "weighted_score": 0.8948 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "classification_multiclass|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1014,19 +1014,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 836, - "total_runtime_ms": 760, - "weighted_score": 0.86 + "p95_latency_ms": 1778, + "total_runtime_ms": 1617, + "weighted_score": 0.7405 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "classification_multiclass|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1034,19 +1034,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 468, - "total_runtime_ms": 426, - "weighted_score": 0.9148 + "p95_latency_ms": 1023, + "total_runtime_ms": 930, + "weighted_score": 0.834 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1054,19 +1054,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1597, - "total_runtime_ms": 1452, - "weighted_score": 0.7635 + "p95_latency_ms": 841, + "total_runtime_ms": 764, + "weighted_score": 0.8598 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "classification_multiclass|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1074,59 +1074,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 903, - "total_runtime_ms": 821, - "weighted_score": 0.8509 + "p95_latency_ms": 467, + "total_runtime_ms": 425, + "weighted_score": 0.9151 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2334, - "total_runtime_ms": 2122, - "weighted_score": 0.6521 + "failure_rate": 0.06, + "p95_latency_ms": 1339, + "total_runtime_ms": 1218, + "weighted_score": 0.7958 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "classification_multiclass|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1292, - "total_runtime_ms": 1175, - "weighted_score": 0.7743 + "failure_rate": 0.03, + "p95_latency_ms": 772, + "total_runtime_ms": 702, + "weighted_score": 0.8706 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:0.5b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1134,19 +1134,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 524, - "total_runtime_ms": 476, - "weighted_score": 0.899 + "p95_latency_ms": 628, + "total_runtime_ms": 571, + "weighted_score": 0.886 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:0.5b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "classification_multiclass|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1154,19 +1154,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 305, - "total_runtime_ms": 278, - "weighted_score": 0.9389 + "p95_latency_ms": 390, + "total_runtime_ms": 355, + "weighted_score": 0.9279 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:1.5b|system_selected|cold", - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1174,19 +1174,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 702, - "total_runtime_ms": 638, - "weighted_score": 0.8768 + "p95_latency_ms": 1015, + "total_runtime_ms": 923, + "weighted_score": 0.8368 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:1.5b|system_selected|warm", - "scenario": "rewrite_constrained", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "classification_multiclass|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1194,19 +1194,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 397, - "total_runtime_ms": 361, - "weighted_score": 0.9253 + "p95_latency_ms": 586, + "total_runtime_ms": 533, + "weighted_score": 0.8977 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:7b|system_selected|cold", - "scenario": "rewrite_constrained", + "combo_id": "classification_multiclass|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1214,19 +1214,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1317, - "total_runtime_ms": 1197, - "weighted_score": 0.7983 + "p95_latency_ms": 558, + "total_runtime_ms": 507, + "weighted_score": 0.8956 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:7b|system_selected|warm", - "scenario": "rewrite_constrained", + "combo_id": "classification_multiclass|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1234,59 +1234,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 774, - "total_runtime_ms": 703, - "weighted_score": 0.8706 + "p95_latency_ms": 326, + "total_runtime_ms": 296, + "weighted_score": 0.9368 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:14b|system_selected|cold", - "scenario": "rewrite_constrained", + "combo_id": "classification_multiclass|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "classification_multiclass", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1944, - "total_runtime_ms": 1767, - "weighted_score": 0.7012 + "failure_rate": 0.06, + "p95_latency_ms": 888, + "total_runtime_ms": 807, + "weighted_score": 0.8538 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "rewrite_constrained|high|qwen2:14b|system_selected|warm", - "scenario": "rewrite_constrained", + "combo_id": "classification_multiclass|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "classification_multiclass", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1089, - "total_runtime_ms": 990, - "weighted_score": 0.8047 + "failure_rate": 0.03, + "p95_latency_ms": 518, + "total_runtime_ms": 471, + "weighted_score": 0.9088 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:0.5b|system_selected|cold", - "scenario": "classification_binary", + "combo_id": "extract_entities|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_entities", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1294,19 +1294,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 983, - "total_runtime_ms": 893, - "weighted_score": 0.8388 + "p95_latency_ms": 994, + "total_runtime_ms": 904, + "weighted_score": 0.8377 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:0.5b|system_selected|warm", - "scenario": "classification_binary", + "combo_id": "extract_entities|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_entities", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1314,19 +1314,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 526, - "total_runtime_ms": 478, - "weighted_score": 0.9025 + "p95_latency_ms": 507, + "total_runtime_ms": 461, + "weighted_score": 0.9044 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:1.5b|system_selected|cold", - "scenario": "classification_binary", + "combo_id": "extract_entities|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_entities", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1334,19 +1334,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1297, - "total_runtime_ms": 1179, - "weighted_score": 0.798 + "p95_latency_ms": 1637, + "total_runtime_ms": 1488, + "weighted_score": 0.7547 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:1.5b|system_selected|warm", - "scenario": "classification_binary", + "combo_id": "extract_entities|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_entities", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1354,19 +1354,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 667, - "total_runtime_ms": 607, - "weighted_score": 0.879 + "p95_latency_ms": 834, + "total_runtime_ms": 758, + "weighted_score": 0.8529 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:7b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "extract_entities|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_entities", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1374,19 +1374,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 2531, - "total_runtime_ms": 2301, - "weighted_score": 0.6414 + "p95_latency_ms": 749, + "total_runtime_ms": 681, + "weighted_score": 0.8689 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:7b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "extract_entities|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_entities", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1394,59 +1394,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1292, - "total_runtime_ms": 1175, - "weighted_score": 0.7802 + "p95_latency_ms": 394, + "total_runtime_ms": 358, + "weighted_score": 0.9225 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:14b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "extract_entities|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_entities", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3722, - "total_runtime_ms": 3383, - "weighted_score": 0.5885 + "failure_rate": 0.06, + "p95_latency_ms": 1227, + "total_runtime_ms": 1115, + "weighted_score": 0.8071 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|low|qwen2:14b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "extract_entities|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_entities", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1895, - "total_runtime_ms": 1722, - "weighted_score": 0.667 + "failure_rate": 0.03, + "p95_latency_ms": 653, + "total_runtime_ms": 594, + "weighted_score": 0.8824 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:0.5b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_entities", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1454,19 +1454,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 747, - "total_runtime_ms": 679, - "weighted_score": 0.8692 + "p95_latency_ms": 569, + "total_runtime_ms": 517, + "weighted_score": 0.892 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:0.5b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "extract_entities|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_entities", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1474,19 +1474,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 384, - "total_runtime_ms": 349, - "weighted_score": 0.9234 + "p95_latency_ms": 338, + "total_runtime_ms": 307, + "weighted_score": 0.9331 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:1.5b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_entities", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1494,19 +1494,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1006, - "total_runtime_ms": 915, - "weighted_score": 0.8362 + "p95_latency_ms": 938, + "total_runtime_ms": 852, + "weighted_score": 0.8445 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:1.5b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "extract_entities|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_entities", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1514,19 +1514,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 541, - "total_runtime_ms": 492, - "weighted_score": 0.9007 + "p95_latency_ms": 519, + "total_runtime_ms": 471, + "weighted_score": 0.9044 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:7b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "extract_entities|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -1534,19 +1534,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1930, - "total_runtime_ms": 1755, - "weighted_score": 0.7166 + "p95_latency_ms": 504, + "total_runtime_ms": 458, + "weighted_score": 0.901 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:7b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "extract_entities|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -1554,699 +1554,699 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 991, - "total_runtime_ms": 901, - "weighted_score": 0.8285 + "p95_latency_ms": 284, + "total_runtime_ms": 258, + "weighted_score": 0.941 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:14b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "extract_entities|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2806, - "total_runtime_ms": 2551, - "weighted_score": 0.6154 + "failure_rate": 0.06, + "p95_latency_ms": 825, + "total_runtime_ms": 750, + "weighted_score": 0.8601 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium|qwen2:14b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "extract_entities|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_entities", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1431, - "total_runtime_ms": 1301, - "weighted_score": 0.7403 + "failure_rate": 0.03, + "p95_latency_ms": 422, + "total_runtime_ms": 384, + "weighted_score": 0.9184 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "extract_strict_json|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.912, "failure_rate": 0.06, - "p95_latency_ms": 599, - "total_runtime_ms": 544, - "weighted_score": 0.889 + "p95_latency_ms": 1126, + "total_runtime_ms": 1024, + "weighted_score": 0.8245 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "extract_strict_json|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.912, "failure_rate": 0.03, - "p95_latency_ms": 313, - "total_runtime_ms": 284, - "weighted_score": 0.9356 + "p95_latency_ms": 683, + "total_runtime_ms": 621, + "weighted_score": 0.8868 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "extract_strict_json|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9008, "failure_rate": 0.06, - "p95_latency_ms": 767, - "total_runtime_ms": 697, - "weighted_score": 0.8669 + "p95_latency_ms": 1883, + "total_runtime_ms": 1712, + "weighted_score": 0.73 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "extract_strict_json|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9008, "failure_rate": 0.03, - "p95_latency_ms": 406, - "total_runtime_ms": 369, - "weighted_score": 0.921 + "p95_latency_ms": 1120, + "total_runtime_ms": 1019, + "weighted_score": 0.8243 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.92, "failure_rate": 0.06, - "p95_latency_ms": 1480, - "total_runtime_ms": 1345, - "weighted_score": 0.7752 + "p95_latency_ms": 873, + "total_runtime_ms": 793, + "weighted_score": 0.8566 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "extract_strict_json|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.92, "failure_rate": 0.03, - "p95_latency_ms": 766, - "total_runtime_ms": 696, - "weighted_score": 0.8646 + "p95_latency_ms": 513, + "total_runtime_ms": 467, + "weighted_score": 0.9105 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2119, - "total_runtime_ms": 1926, - "weighted_score": 0.6736 + "task_success_rate": 0.94, + "schema_validity_rate": 0.9088, + "failure_rate": 0.06, + "p95_latency_ms": 1437, + "total_runtime_ms": 1306, + "weighted_score": 0.786 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "extract_strict_json|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1111, - "total_runtime_ms": 1010, - "weighted_score": 0.7924 + "task_success_rate": 0.97, + "schema_validity_rate": 0.9088, + "failure_rate": 0.03, + "p95_latency_ms": 842, + "total_runtime_ms": 765, + "weighted_score": 0.8636 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:0.5b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.926, "failure_rate": 0.06, - "p95_latency_ms": 497, - "total_runtime_ms": 452, - "weighted_score": 0.9017 + "p95_latency_ms": 669, + "total_runtime_ms": 608, + "weighted_score": 0.882 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:0.5b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "extract_strict_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.926, "failure_rate": 0.03, - "p95_latency_ms": 296, - "total_runtime_ms": 269, - "weighted_score": 0.9399 + "p95_latency_ms": 421, + "total_runtime_ms": 383, + "weighted_score": 0.9247 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:1.5b|system_selected|cold", - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9148, "failure_rate": 0.06, - "p95_latency_ms": 627, - "total_runtime_ms": 570, - "weighted_score": 0.8842 + "p95_latency_ms": 1084, + "total_runtime_ms": 985, + "weighted_score": 0.8299 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:1.5b|system_selected|warm", - "scenario": "classification_binary", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "extract_strict_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9148, "failure_rate": 0.03, - "p95_latency_ms": 357, - "total_runtime_ms": 324, - "weighted_score": 0.9293 + "p95_latency_ms": 665, + "total_runtime_ms": 604, + "weighted_score": 0.8898 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:7b|system_selected|cold", - "scenario": "classification_binary", + "combo_id": "extract_strict_json|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.929, "failure_rate": 0.06, - "p95_latency_ms": 1228, - "total_runtime_ms": 1116, - "weighted_score": 0.8072 + "p95_latency_ms": 558, + "total_runtime_ms": 507, + "weighted_score": 0.8956 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:7b|system_selected|warm", - "scenario": "classification_binary", + "combo_id": "extract_strict_json|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.929, "failure_rate": 0.03, - "p95_latency_ms": 635, - "total_runtime_ms": 577, - "weighted_score": 0.8845 + "p95_latency_ms": 350, + "total_runtime_ms": 318, + "weighted_score": 0.9344 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:14b|system_selected|cold", - "scenario": "classification_binary", + "combo_id": "extract_strict_json|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1784, - "total_runtime_ms": 1621, - "weighted_score": 0.7173 + "task_success_rate": 0.94, + "schema_validity_rate": 0.9178, + "failure_rate": 0.06, + "p95_latency_ms": 929, + "total_runtime_ms": 845, + "weighted_score": 0.8497 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_binary|high|qwen2:14b|system_selected|warm", - "scenario": "classification_binary", + "combo_id": "extract_strict_json|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_strict_json", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 933, - "total_runtime_ms": 848, - "weighted_score": 0.8203 + "task_success_rate": 0.97, + "schema_validity_rate": 0.9178, + "failure_rate": 0.03, + "p95_latency_ms": 548, + "total_runtime_ms": 499, + "weighted_score": 0.9057 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:0.5b|system_selected|cold", - "scenario": "classification_multiclass", + "combo_id": "extract_nested_json|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.912, "failure_rate": 0.06, - "p95_latency_ms": 1081, - "total_runtime_ms": 983, - "weighted_score": 0.829 + "p95_latency_ms": 1334, + "total_runtime_ms": 1213, + "weighted_score": 0.8037 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:0.5b|system_selected|warm", - "scenario": "classification_multiclass", + "combo_id": "extract_nested_json|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.912, "failure_rate": 0.03, - "p95_latency_ms": 607, - "total_runtime_ms": 552, - "weighted_score": 0.8944 + "p95_latency_ms": 884, + "total_runtime_ms": 804, + "weighted_score": 0.8667 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:1.5b|system_selected|cold", - "scenario": "classification_multiclass", + "combo_id": "extract_nested_json|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9008, "failure_rate": 0.06, - "p95_latency_ms": 1413, - "total_runtime_ms": 1284, - "weighted_score": 0.7864 + "p95_latency_ms": 2240, + "total_runtime_ms": 2036, + "weighted_score": 0.6943 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:1.5b|system_selected|warm", - "scenario": "classification_multiclass", + "combo_id": "extract_nested_json|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9008, "failure_rate": 0.03, - "p95_latency_ms": 817, - "total_runtime_ms": 743, - "weighted_score": 0.864 + "p95_latency_ms": 1479, + "total_runtime_ms": 1345, + "weighted_score": 0.7884 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:7b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.92, "failure_rate": 0.06, - "p95_latency_ms": 2807, - "total_runtime_ms": 2552, - "weighted_score": 0.6414 + "p95_latency_ms": 1004, + "total_runtime_ms": 913, + "weighted_score": 0.8434 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:7b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "extract_nested_json|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.92, "failure_rate": 0.03, - "p95_latency_ms": 1543, - "total_runtime_ms": 1403, - "weighted_score": 0.7551 + "p95_latency_ms": 679, + "total_runtime_ms": 617, + "weighted_score": 0.8939 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:14b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 4107, - "total_runtime_ms": 3734, - "weighted_score": 0.5885 + "task_success_rate": 0.94, + "schema_validity_rate": 0.9088, + "failure_rate": 0.06, + "p95_latency_ms": 1702, + "total_runtime_ms": 1547, + "weighted_score": 0.7596 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|low|qwen2:14b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "extract_nested_json|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2251, - "total_runtime_ms": 2047, - "weighted_score": 0.6313 + "task_success_rate": 0.97, + "schema_validity_rate": 0.9088, + "failure_rate": 0.03, + "p95_latency_ms": 1134, + "total_runtime_ms": 1031, + "weighted_score": 0.8344 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:0.5b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.926, "failure_rate": 0.06, - "p95_latency_ms": 806, - "total_runtime_ms": 733, - "weighted_score": 0.8632 + "p95_latency_ms": 800, + "total_runtime_ms": 727, + "weighted_score": 0.8689 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:0.5b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "extract_nested_json|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.926, "failure_rate": 0.03, - "p95_latency_ms": 475, - "total_runtime_ms": 432, - "weighted_score": 0.9143 + "p95_latency_ms": 523, + "total_runtime_ms": 475, + "weighted_score": 0.9146 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:1.5b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9148, "failure_rate": 0.06, - "p95_latency_ms": 1077, - "total_runtime_ms": 979, - "weighted_score": 0.8291 + "p95_latency_ms": 1300, + "total_runtime_ms": 1182, + "weighted_score": 0.8083 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:1.5b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "extract_nested_json|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.9148, "failure_rate": 0.03, - "p95_latency_ms": 616, - "total_runtime_ms": 560, - "weighted_score": 0.8932 + "p95_latency_ms": 846, + "total_runtime_ms": 769, + "weighted_score": 0.8717 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:7b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "extract_nested_json|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { "task_success_rate": 0.94, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.929, "failure_rate": 0.06, - "p95_latency_ms": 2090, - "total_runtime_ms": 1900, - "weighted_score": 0.7006 + "p95_latency_ms": 679, + "total_runtime_ms": 617, + "weighted_score": 0.8835 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:7b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "extract_nested_json|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { "task_success_rate": 0.97, - "schema_validity_rate": 0.98, + "schema_validity_rate": 0.929, "failure_rate": 0.03, - "p95_latency_ms": 1176, - "total_runtime_ms": 1069, - "weighted_score": 0.81 + "p95_latency_ms": 460, + "total_runtime_ms": 418, + "weighted_score": 0.9234 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:14b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "extract_nested_json|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3092, - "total_runtime_ms": 2811, - "weighted_score": 0.6154 + "task_success_rate": 0.94, + "schema_validity_rate": 0.9178, + "failure_rate": 0.06, + "p95_latency_ms": 1079, + "total_runtime_ms": 981, + "weighted_score": 0.8347 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium|qwen2:14b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "extract_nested_json|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "extract_nested_json", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1700, - "total_runtime_ms": 1546, - "weighted_score": 0.7133 + "task_success_rate": 0.97, + "schema_validity_rate": 0.9178, + "failure_rate": 0.03, + "p95_latency_ms": 732, + "total_runtime_ms": 666, + "weighted_score": 0.8874 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "summarize_contextual|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2254,19 +2254,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 639, - "total_runtime_ms": 581, - "weighted_score": 0.885 + "p95_latency_ms": 996, + "total_runtime_ms": 906, + "weighted_score": 0.8375 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "summarize_contextual|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2274,19 +2274,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 366, - "total_runtime_ms": 333, - "weighted_score": 0.9303 + "p95_latency_ms": 529, + "total_runtime_ms": 481, + "weighted_score": 0.9023 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "summarize_contextual|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2294,19 +2294,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 827, - "total_runtime_ms": 752, - "weighted_score": 0.8609 + "p95_latency_ms": 1621, + "total_runtime_ms": 1473, + "weighted_score": 0.7562 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "summarize_contextual|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2314,19 +2314,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 497, - "total_runtime_ms": 452, - "weighted_score": 0.9119 + "p95_latency_ms": 856, + "total_runtime_ms": 778, + "weighted_score": 0.8507 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2334,19 +2334,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1586, - "total_runtime_ms": 1441, - "weighted_score": 0.7646 + "p95_latency_ms": 747, + "total_runtime_ms": 679, + "weighted_score": 0.8691 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "summarize_contextual|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2354,59 +2354,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 906, - "total_runtime_ms": 823, - "weighted_score": 0.8506 + "p95_latency_ms": 422, + "total_runtime_ms": 384, + "weighted_score": 0.9196 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2333, - "total_runtime_ms": 2121, - "weighted_score": 0.6523 + "failure_rate": 0.06, + "p95_latency_ms": 1245, + "total_runtime_ms": 1132, + "weighted_score": 0.8053 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "summarize_contextual|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1306, - "total_runtime_ms": 1187, - "weighted_score": 0.7729 + "failure_rate": 0.03, + "p95_latency_ms": 651, + "total_runtime_ms": 592, + "weighted_score": 0.8826 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:0.5b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2414,19 +2414,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 538, - "total_runtime_ms": 489, - "weighted_score": 0.8976 + "p95_latency_ms": 606, + "total_runtime_ms": 552, + "weighted_score": 0.8882 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:0.5b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "summarize_contextual|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2434,19 +2434,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 326, - "total_runtime_ms": 296, - "weighted_score": 0.9368 + "p95_latency_ms": 329, + "total_runtime_ms": 299, + "weighted_score": 0.934 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:1.5b|system_selected|cold", - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2454,19 +2454,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 692, - "total_runtime_ms": 629, - "weighted_score": 0.8778 + "p95_latency_ms": 929, + "total_runtime_ms": 845, + "weighted_score": 0.8454 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:1.5b|system_selected|warm", - "scenario": "classification_multiclass", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "summarize_contextual|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2474,19 +2474,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 422, - "total_runtime_ms": 383, - "weighted_score": 0.9228 + "p95_latency_ms": 506, + "total_runtime_ms": 460, + "weighted_score": 0.9057 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:7b|system_selected|cold", - "scenario": "classification_multiclass", + "combo_id": "summarize_contextual|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2494,19 +2494,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1365, - "total_runtime_ms": 1241, - "weighted_score": 0.7935 + "p95_latency_ms": 512, + "total_runtime_ms": 465, + "weighted_score": 0.9002 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:7b|system_selected|warm", - "scenario": "classification_multiclass", + "combo_id": "summarize_contextual|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2514,59 +2514,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 764, - "total_runtime_ms": 694, - "weighted_score": 0.8716 + "p95_latency_ms": 274, + "total_runtime_ms": 249, + "weighted_score": 0.942 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:14b|system_selected|cold", - "scenario": "classification_multiclass", + "combo_id": "summarize_contextual|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1934, - "total_runtime_ms": 1758, - "weighted_score": 0.7022 + "failure_rate": 0.06, + "p95_latency_ms": 812, + "total_runtime_ms": 738, + "weighted_score": 0.8614 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "classification_multiclass|high|qwen2:14b|system_selected|warm", - "scenario": "classification_multiclass", + "combo_id": "summarize_contextual|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "summarize_contextual", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1091, - "total_runtime_ms": 991, - "weighted_score": 0.8045 + "failure_rate": 0.03, + "p95_latency_ms": 414, + "total_runtime_ms": 377, + "weighted_score": 0.9192 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:0.5b|system_selected|cold", - "scenario": "extract_entities", + "combo_id": "analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2574,19 +2574,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 986, - "total_runtime_ms": 896, - "weighted_score": 0.8385 + "p95_latency_ms": 1218, + "total_runtime_ms": 1107, + "weighted_score": 0.8154 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:0.5b|system_selected|warm", - "scenario": "extract_entities", + "combo_id": "analyze_reasoning|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2594,19 +2594,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 511, - "total_runtime_ms": 465, - "weighted_score": 0.904 + "p95_latency_ms": 775, + "total_runtime_ms": 705, + "weighted_score": 0.8776 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:1.5b|system_selected|cold", - "scenario": "extract_entities", + "combo_id": "analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2614,19 +2614,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1312, - "total_runtime_ms": 1193, - "weighted_score": 0.7965 + "p95_latency_ms": 2024, + "total_runtime_ms": 1840, + "weighted_score": 0.7159 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:1.5b|system_selected|warm", - "scenario": "extract_entities", + "combo_id": "analyze_reasoning|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2634,19 +2634,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 702, - "total_runtime_ms": 638, - "weighted_score": 0.8755 + "p95_latency_ms": 1210, + "total_runtime_ms": 1100, + "weighted_score": 0.8154 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:7b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2654,19 +2654,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 2534, - "total_runtime_ms": 2303, - "weighted_score": 0.6414 + "p95_latency_ms": 933, + "total_runtime_ms": 848, + "weighted_score": 0.8506 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:7b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "analyze_reasoning|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2674,59 +2674,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1291, - "total_runtime_ms": 1173, - "weighted_score": 0.7804 + "p95_latency_ms": 591, + "total_runtime_ms": 537, + "weighted_score": 0.9027 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:14b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3696, - "total_runtime_ms": 3360, - "weighted_score": 0.5885 + "failure_rate": 0.06, + "p95_latency_ms": 1534, + "total_runtime_ms": 1394, + "weighted_score": 0.7764 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|low|qwen2:14b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "analyze_reasoning|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1890, - "total_runtime_ms": 1718, - "weighted_score": 0.6675 + "failure_rate": 0.03, + "p95_latency_ms": 938, + "total_runtime_ms": 852, + "weighted_score": 0.854 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:0.5b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2734,19 +2734,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 747, - "total_runtime_ms": 679, - "weighted_score": 0.8691 + "p95_latency_ms": 729, + "total_runtime_ms": 662, + "weighted_score": 0.876 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:0.5b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "analyze_reasoning|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -2754,219 +2754,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 401, - "total_runtime_ms": 364, - "weighted_score": 0.9217 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:1.5b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 994, - "total_runtime_ms": 903, - "weighted_score": 0.8374 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:1.5b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 518, - "total_runtime_ms": 471, - "weighted_score": 0.903 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:7b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1927, - "total_runtime_ms": 1752, - "weighted_score": 0.7168 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:7b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 990, - "total_runtime_ms": 900, - "weighted_score": 0.8286 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:14b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2815, - "total_runtime_ms": 2559, - "weighted_score": 0.6154 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium|qwen2:14b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1409, - "total_runtime_ms": 1281, - "weighted_score": 0.7425 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 582, - "total_runtime_ms": 529, - "weighted_score": 0.8906 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 331, - "total_runtime_ms": 301, - "weighted_score": 0.9338 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 770, - "total_runtime_ms": 700, - "weighted_score": 0.8666 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 401, - "total_runtime_ms": 365, - "weighted_score": 0.9215 + "p95_latency_ms": 487, + "total_runtime_ms": 443, + "weighted_score": 0.9182 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "extract_entities", + "combo_id": "analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -2974,139 +2774,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1444, - "total_runtime_ms": 1313, - "weighted_score": 0.7788 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 775, - "total_runtime_ms": 704, - "weighted_score": 0.8637 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2129, - "total_runtime_ms": 1936, - "weighted_score": 0.6726 + "p95_latency_ms": 1172, + "total_runtime_ms": 1065, + "weighted_score": 0.8211 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "extract_entities", + "combo_id": "analyze_reasoning|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1114, - "total_runtime_ms": 1013, - "weighted_score": 0.7921 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:0.5b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 493, - "total_runtime_ms": 448, - "weighted_score": 0.9021 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:0.5b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 282, - "total_runtime_ms": 257, - "weighted_score": 0.9412 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:1.5b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 642, - "total_runtime_ms": 583, - "weighted_score": 0.8828 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:1.5b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -3114,19 +2794,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 344, - "total_runtime_ms": 313, - "weighted_score": 0.9306 + "p95_latency_ms": 704, + "total_runtime_ms": 640, + "weighted_score": 0.8859 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:7b|system_selected|cold", - "scenario": "extract_entities", + "combo_id": "analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -3134,4239 +2814,79 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1234, - "total_runtime_ms": 1122, - "weighted_score": 0.8066 + "p95_latency_ms": 598, + "total_runtime_ms": 543, + "weighted_score": 0.8916 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:7b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 632, - "total_runtime_ms": 574, - "weighted_score": 0.8848 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:14b|system_selected|cold", - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1794, - "total_runtime_ms": 1631, - "weighted_score": 0.7162 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_entities|high|qwen2:14b|system_selected|warm", - "scenario": "extract_entities", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 940, - "total_runtime_ms": 854, - "weighted_score": 0.8196 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:0.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.912, - "failure_rate": 0.06, - "p95_latency_ms": 1152, - "total_runtime_ms": 1047, - "weighted_score": 0.822 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:0.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.912, - "failure_rate": 0.03, - "p95_latency_ms": 682, - "total_runtime_ms": 620, - "weighted_score": 0.8869 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:1.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9064, - "failure_rate": 0.06, - "p95_latency_ms": 1515, - "total_runtime_ms": 1377, - "weighted_score": 0.7762 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:1.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9064, - "failure_rate": 0.03, - "p95_latency_ms": 888, - "total_runtime_ms": 807, - "weighted_score": 0.8569 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:7b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.8848, - "failure_rate": 0.06, - "p95_latency_ms": 2980, - "total_runtime_ms": 2709, - "weighted_score": 0.6414 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:7b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.8848, - "failure_rate": 0.03, - "p95_latency_ms": 1747, - "total_runtime_ms": 1588, - "weighted_score": 0.7347 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:14b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.864, - "failure_rate": 0.09, - "p95_latency_ms": 4383, - "total_runtime_ms": 3985, - "weighted_score": 0.5885 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|low|qwen2:14b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.864, - "failure_rate": 0.06, - "p95_latency_ms": 2503, - "total_runtime_ms": 2276, - "weighted_score": 0.6065 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:0.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.92, - "failure_rate": 0.06, - "p95_latency_ms": 884, - "total_runtime_ms": 803, - "weighted_score": 0.8555 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:0.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.92, - "failure_rate": 0.03, - "p95_latency_ms": 516, - "total_runtime_ms": 469, - "weighted_score": 0.9102 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:1.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9144, - "failure_rate": 0.06, - "p95_latency_ms": 1156, - "total_runtime_ms": 1051, - "weighted_score": 0.8212 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:1.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9144, - "failure_rate": 0.03, - "p95_latency_ms": 692, - "total_runtime_ms": 630, - "weighted_score": 0.8856 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:7b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.8928, - "failure_rate": 0.06, - "p95_latency_ms": 2242, - "total_runtime_ms": 2038, - "weighted_score": 0.6854 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:7b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.8928, - "failure_rate": 0.03, - "p95_latency_ms": 1313, - "total_runtime_ms": 1193, - "weighted_score": 0.7963 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:14b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.872, - "failure_rate": 0.09, - "p95_latency_ms": 3273, - "total_runtime_ms": 2976, - "weighted_score": 0.6154 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium|qwen2:14b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.872, - "failure_rate": 0.06, - "p95_latency_ms": 1913, - "total_runtime_ms": 1739, - "weighted_score": 0.6921 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.926, - "failure_rate": 0.06, - "p95_latency_ms": 673, - "total_runtime_ms": 612, - "weighted_score": 0.8816 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.926, - "failure_rate": 0.03, - "p95_latency_ms": 414, - "total_runtime_ms": 376, - "weighted_score": 0.9255 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9204, - "failure_rate": 0.06, - "p95_latency_ms": 884, - "total_runtime_ms": 804, - "weighted_score": 0.8552 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9204, - "failure_rate": 0.03, - "p95_latency_ms": 538, - "total_runtime_ms": 489, - "weighted_score": 0.9078 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.8988, - "failure_rate": 0.06, - "p95_latency_ms": 1695, - "total_runtime_ms": 1541, - "weighted_score": 0.7537 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.8988, - "failure_rate": 0.03, - "p95_latency_ms": 1021, - "total_runtime_ms": 928, - "weighted_score": 0.8391 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.878, - "failure_rate": 0.09, - "p95_latency_ms": 2487, - "total_runtime_ms": 2261, - "weighted_score": 0.6368 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.878, - "failure_rate": 0.06, - "p95_latency_ms": 1449, - "total_runtime_ms": 1317, - "weighted_score": 0.7586 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:0.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.929, - "failure_rate": 0.06, - "p95_latency_ms": 568, - "total_runtime_ms": 516, - "weighted_score": 0.8946 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:0.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.929, - "failure_rate": 0.03, - "p95_latency_ms": 384, - "total_runtime_ms": 349, - "weighted_score": 0.931 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:1.5b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9234, - "failure_rate": 0.06, - "p95_latency_ms": 753, - "total_runtime_ms": 685, - "weighted_score": 0.8717 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:1.5b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9234, - "failure_rate": 0.03, - "p95_latency_ms": 447, - "total_runtime_ms": 406, - "weighted_score": 0.9203 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:7b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9018, - "failure_rate": 0.06, - "p95_latency_ms": 1435, - "total_runtime_ms": 1305, - "weighted_score": 0.7865 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:7b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9018, - "failure_rate": 0.03, - "p95_latency_ms": 855, - "total_runtime_ms": 777, - "weighted_score": 0.8625 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:14b|system_selected|cold", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.881, - "failure_rate": 0.09, - "p95_latency_ms": 2086, - "total_runtime_ms": 1897, - "weighted_score": 0.687 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_strict_json|high|qwen2:14b|system_selected|warm", - "scenario": "extract_strict_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.881, - "failure_rate": 0.06, - "p95_latency_ms": 1229, - "total_runtime_ms": 1117, - "weighted_score": 0.7907 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:0.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.912, - "failure_rate": 0.06, - "p95_latency_ms": 1348, - "total_runtime_ms": 1225, - "weighted_score": 0.8023 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:0.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.912, - "failure_rate": 0.03, - "p95_latency_ms": 899, - "total_runtime_ms": 818, - "weighted_score": 0.8652 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:1.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9064, - "failure_rate": 0.06, - "p95_latency_ms": 1803, - "total_runtime_ms": 1639, - "weighted_score": 0.7474 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:1.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9064, - "failure_rate": 0.03, - "p95_latency_ms": 1205, - "total_runtime_ms": 1096, - "weighted_score": 0.8252 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:7b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.8848, - "failure_rate": 0.06, - "p95_latency_ms": 3513, - "total_runtime_ms": 3194, - "weighted_score": 0.6414 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:7b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.8848, - "failure_rate": 0.03, - "p95_latency_ms": 2291, - "total_runtime_ms": 2083, - "weighted_score": 0.6804 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:14b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.864, - "failure_rate": 0.09, - "p95_latency_ms": 5185, - "total_runtime_ms": 4714, - "weighted_score": 0.5885 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|low|qwen2:14b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.864, - "failure_rate": 0.06, - "p95_latency_ms": 3346, - "total_runtime_ms": 3042, - "weighted_score": 0.6065 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:0.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.92, - "failure_rate": 0.06, - "p95_latency_ms": 1037, - "total_runtime_ms": 942, - "weighted_score": 0.8402 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:0.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.92, - "failure_rate": 0.03, - "p95_latency_ms": 671, - "total_runtime_ms": 610, - "weighted_score": 0.8948 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:1.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9144, - "failure_rate": 0.06, - "p95_latency_ms": 1377, - "total_runtime_ms": 1252, - "weighted_score": 0.7991 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:1.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9144, - "failure_rate": 0.03, - "p95_latency_ms": 908, - "total_runtime_ms": 825, - "weighted_score": 0.864 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:7b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.8928, - "failure_rate": 0.06, - "p95_latency_ms": 2659, - "total_runtime_ms": 2417, - "weighted_score": 0.6596 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:7b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.8928, - "failure_rate": 0.03, - "p95_latency_ms": 1719, - "total_runtime_ms": 1562, - "weighted_score": 0.7557 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:14b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.872, - "failure_rate": 0.09, - "p95_latency_ms": 3912, - "total_runtime_ms": 3556, - "weighted_score": 0.6154 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium|qwen2:14b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.872, - "failure_rate": 0.06, - "p95_latency_ms": 2541, - "total_runtime_ms": 2310, - "weighted_score": 0.6334 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.926, - "failure_rate": 0.06, - "p95_latency_ms": 765, - "total_runtime_ms": 696, - "weighted_score": 0.8723 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.926, - "failure_rate": 0.03, - "p95_latency_ms": 522, - "total_runtime_ms": 475, - "weighted_score": 0.9147 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9204, - "failure_rate": 0.06, - "p95_latency_ms": 1036, - "total_runtime_ms": 942, - "weighted_score": 0.84 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9204, - "failure_rate": 0.03, - "p95_latency_ms": 694, - "total_runtime_ms": 631, - "weighted_score": 0.8922 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.8988, - "failure_rate": 0.06, - "p95_latency_ms": 2007, - "total_runtime_ms": 1825, - "weighted_score": 0.7224 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.8988, - "failure_rate": 0.03, - "p95_latency_ms": 1308, - "total_runtime_ms": 1189, - "weighted_score": 0.8104 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.878, - "failure_rate": 0.09, - "p95_latency_ms": 2964, - "total_runtime_ms": 2694, - "weighted_score": 0.6355 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.878, - "failure_rate": 0.06, - "p95_latency_ms": 1921, - "total_runtime_ms": 1746, - "weighted_score": 0.7114 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:0.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.929, - "failure_rate": 0.06, - "p95_latency_ms": 681, - "total_runtime_ms": 619, - "weighted_score": 0.8833 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:0.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.929, - "failure_rate": 0.03, - "p95_latency_ms": 455, - "total_runtime_ms": 413, - "weighted_score": 0.9239 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:1.5b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9234, - "failure_rate": 0.06, - "p95_latency_ms": 886, - "total_runtime_ms": 806, - "weighted_score": 0.8583 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:1.5b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9234, - "failure_rate": 0.03, - "p95_latency_ms": 596, - "total_runtime_ms": 542, - "weighted_score": 0.9054 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:7b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.9018, - "failure_rate": 0.06, - "p95_latency_ms": 1678, - "total_runtime_ms": 1526, - "weighted_score": 0.7621 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:7b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.9018, - "failure_rate": 0.03, - "p95_latency_ms": 1108, - "total_runtime_ms": 1007, - "weighted_score": 0.8372 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:14b|system_selected|cold", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.881, - "failure_rate": 0.09, - "p95_latency_ms": 2448, - "total_runtime_ms": 2225, - "weighted_score": 0.6509 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "extract_nested_json|high|qwen2:14b|system_selected|warm", - "scenario": "extract_nested_json", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.881, - "failure_rate": 0.06, - "p95_latency_ms": 1602, - "total_runtime_ms": 1456, - "weighted_score": 0.7534 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:0.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 979, - "total_runtime_ms": 890, - "weighted_score": 0.8392 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:0.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 550, - "total_runtime_ms": 500, - "weighted_score": 0.9001 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:1.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1317, - "total_runtime_ms": 1197, - "weighted_score": 0.796 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:1.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 692, - "total_runtime_ms": 629, - "weighted_score": 0.8765 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:7b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2549, - "total_runtime_ms": 2318, - "weighted_score": 0.6414 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:7b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1312, - "total_runtime_ms": 1193, - "weighted_score": 0.7782 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:14b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3718, - "total_runtime_ms": 3380, - "weighted_score": 0.5885 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|low|qwen2:14b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1886, - "total_runtime_ms": 1714, - "weighted_score": 0.6679 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:0.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 752, - "total_runtime_ms": 684, - "weighted_score": 0.8686 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:0.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 413, - "total_runtime_ms": 376, - "weighted_score": 0.9205 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:1.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 992, - "total_runtime_ms": 901, - "weighted_score": 0.8376 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:1.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 526, - "total_runtime_ms": 478, - "weighted_score": 0.9022 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:7b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1930, - "total_runtime_ms": 1755, - "weighted_score": 0.7166 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:7b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 976, - "total_runtime_ms": 887, - "weighted_score": 0.83 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:14b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2816, - "total_runtime_ms": 2560, - "weighted_score": 0.6154 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium|qwen2:14b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1432, - "total_runtime_ms": 1302, - "weighted_score": 0.7402 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 581, - "total_runtime_ms": 528, - "weighted_score": 0.8908 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 312, - "total_runtime_ms": 284, - "weighted_score": 0.9357 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 764, - "total_runtime_ms": 695, - "weighted_score": 0.8672 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 407, - "total_runtime_ms": 370, - "weighted_score": 0.9209 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1454, - "total_runtime_ms": 1322, - "weighted_score": 0.7778 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 757, - "total_runtime_ms": 688, - "weighted_score": 0.8655 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2124, - "total_runtime_ms": 1930, - "weighted_score": 0.6732 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1113, - "total_runtime_ms": 1012, - "weighted_score": 0.7922 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:0.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 486, - "total_runtime_ms": 441, - "weighted_score": 0.9028 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:0.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 276, - "total_runtime_ms": 251, - "weighted_score": 0.9418 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:1.5b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 668, - "total_runtime_ms": 607, - "weighted_score": 0.8802 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:1.5b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 356, - "total_runtime_ms": 324, - "weighted_score": 0.9294 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:7b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1230, - "total_runtime_ms": 1118, - "weighted_score": 0.807 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:7b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 648, - "total_runtime_ms": 589, - "weighted_score": 0.8832 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:14b|system_selected|cold", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1762, - "total_runtime_ms": 1602, - "weighted_score": 0.7194 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "summarize_contextual|high|qwen2:14b|system_selected|warm", - "scenario": "summarize_contextual", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 950, - "total_runtime_ms": 864, - "weighted_score": 0.8186 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:0.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1211, - "total_runtime_ms": 1101, - "weighted_score": 0.816 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:0.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 740, - "total_runtime_ms": 673, - "weighted_score": 0.8811 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:1.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1601, - "total_runtime_ms": 1456, - "weighted_score": 0.7676 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:1.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 988, - "total_runtime_ms": 898, - "weighted_score": 0.8469 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:7b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 3138, - "total_runtime_ms": 2853, - "weighted_score": 0.6414 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:7b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1935, - "total_runtime_ms": 1759, - "weighted_score": 0.7159 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:14b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 4629, - "total_runtime_ms": 4208, - "weighted_score": 0.5885 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|low|qwen2:14b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2814, - "total_runtime_ms": 2558, - "weighted_score": 0.6065 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:0.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 940, - "total_runtime_ms": 855, - "weighted_score": 0.8498 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:0.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 591, - "total_runtime_ms": 537, - "weighted_score": 0.9028 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:1.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1205, - "total_runtime_ms": 1095, - "weighted_score": 0.8163 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:1.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 757, - "total_runtime_ms": 689, - "weighted_score": 0.8791 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:7b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2380, - "total_runtime_ms": 2164, - "weighted_score": 0.6716 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:7b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1447, - "total_runtime_ms": 1315, - "weighted_score": 0.7829 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:14b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3494, - "total_runtime_ms": 3176, - "weighted_score": 0.6154 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium|qwen2:14b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2106, - "total_runtime_ms": 1915, - "weighted_score": 0.6727 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 699, - "total_runtime_ms": 636, - "weighted_score": 0.8789 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 468, - "total_runtime_ms": 426, - "weighted_score": 0.92 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 932, - "total_runtime_ms": 847, - "weighted_score": 0.8504 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 612, - "total_runtime_ms": 556, - "weighted_score": 0.9004 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1797, - "total_runtime_ms": 1634, - "weighted_score": 0.7434 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1112, - "total_runtime_ms": 1011, - "weighted_score": 0.83 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2659, - "total_runtime_ms": 2417, - "weighted_score": 0.6355 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1606, - "total_runtime_ms": 1460, - "weighted_score": 0.7429 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:0.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 600, - "total_runtime_ms": 546, - "weighted_score": 0.8914 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:0.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 389, - "total_runtime_ms": 354, - "weighted_score": 0.9305 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:1.5b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 786, - "total_runtime_ms": 715, - "weighted_score": 0.8684 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:1.5b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 490, - "total_runtime_ms": 445, - "weighted_score": 0.916 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:7b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1536, - "total_runtime_ms": 1396, - "weighted_score": 0.7764 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:7b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 961, - "total_runtime_ms": 874, - "weighted_score": 0.8519 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:14b|system_selected|cold", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2211, - "total_runtime_ms": 2010, - "weighted_score": 0.6745 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "analyze_reasoning|high|qwen2:14b|system_selected|warm", - "scenario": "analyze_reasoning", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1360, - "total_runtime_ms": 1236, - "weighted_score": 0.7776 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:0.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1214, - "total_runtime_ms": 1104, - "weighted_score": 0.8157 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:0.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 760, - "total_runtime_ms": 691, - "weighted_score": 0.8791 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:1.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1626, - "total_runtime_ms": 1478, - "weighted_score": 0.7651 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:1.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1016, - "total_runtime_ms": 924, - "weighted_score": 0.8441 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:7b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 3172, - "total_runtime_ms": 2884, - "weighted_score": 0.6414 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:7b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1907, - "total_runtime_ms": 1734, - "weighted_score": 0.7187 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:14b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 4637, - "total_runtime_ms": 4216, - "weighted_score": 0.5885 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|low|qwen2:14b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2824, - "total_runtime_ms": 2567, - "weighted_score": 0.6065 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:0.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 947, - "total_runtime_ms": 861, - "weighted_score": 0.8491 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:0.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 582, - "total_runtime_ms": 530, - "weighted_score": 0.9036 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:1.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1221, - "total_runtime_ms": 1110, - "weighted_score": 0.8147 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:1.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 784, - "total_runtime_ms": 712, - "weighted_score": 0.8764 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:7b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2381, - "total_runtime_ms": 2164, - "weighted_score": 0.6715 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:7b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1445, - "total_runtime_ms": 1314, - "weighted_score": 0.783 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:14b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3495, - "total_runtime_ms": 3177, - "weighted_score": 0.6154 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium|qwen2:14b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2117, - "total_runtime_ms": 1924, - "weighted_score": 0.6717 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 723, - "total_runtime_ms": 657, - "weighted_score": 0.8766 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 448, - "total_runtime_ms": 408, - "weighted_score": 0.922 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 928, - "total_runtime_ms": 843, - "weighted_score": 0.8508 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 606, - "total_runtime_ms": 551, - "weighted_score": 0.901 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1804, - "total_runtime_ms": 1640, - "weighted_score": 0.7428 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1103, - "total_runtime_ms": 1003, - "weighted_score": 0.8309 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2635, - "total_runtime_ms": 2395, - "weighted_score": 0.6355 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1612, - "total_runtime_ms": 1466, - "weighted_score": 0.7423 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:0.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 623, - "total_runtime_ms": 566, - "weighted_score": 0.8891 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:0.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 398, - "total_runtime_ms": 361, - "weighted_score": 0.9296 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:1.5b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 781, - "total_runtime_ms": 710, - "weighted_score": 0.8689 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:1.5b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 495, - "total_runtime_ms": 450, - "weighted_score": 0.9155 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:7b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1517, - "total_runtime_ms": 1379, - "weighted_score": 0.7783 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:7b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 928, - "total_runtime_ms": 843, - "weighted_score": 0.8552 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:14b|system_selected|cold", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2227, - "total_runtime_ms": 2025, - "weighted_score": 0.6729 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "compare_structured|high|qwen2:14b|system_selected|warm", - "scenario": "compare_structured", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1349, - "total_runtime_ms": 1227, - "weighted_score": 0.7787 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:0.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 967, - "total_runtime_ms": 879, - "weighted_score": 0.8404 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:0.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.873, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 516, - "total_runtime_ms": 469, - "weighted_score": 0.8648 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:1.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.752, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1313, - "total_runtime_ms": 1194, - "weighted_score": 0.7212 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:1.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 685, - "total_runtime_ms": 623, - "weighted_score": 0.8772 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:7b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.752, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2519, - "total_runtime_ms": 2290, - "weighted_score": 0.5662 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:7b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.873, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1286, - "total_runtime_ms": 1169, - "weighted_score": 0.742 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:14b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.728, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3744, - "total_runtime_ms": 3404, - "weighted_score": 0.5157 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|low|qwen2:14b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1892, - "total_runtime_ms": 1720, - "weighted_score": 0.6297 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:0.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.752, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 753, - "total_runtime_ms": 685, - "weighted_score": 0.7933 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:0.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.873, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 410, - "total_runtime_ms": 373, - "weighted_score": 0.882 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:1.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 993, - "total_runtime_ms": 903, - "weighted_score": 0.7999 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:1.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.873, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 522, - "total_runtime_ms": 474, - "weighted_score": 0.8638 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:7b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1929, - "total_runtime_ms": 1753, - "weighted_score": 0.7167 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:7b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 998, - "total_runtime_ms": 907, - "weighted_score": 0.8278 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:14b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.728, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2804, - "total_runtime_ms": 2549, - "weighted_score": 0.5426 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium|qwen2:14b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1424, - "total_runtime_ms": 1294, - "weighted_score": 0.7034 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 596, - "total_runtime_ms": 541, - "weighted_score": 0.8517 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 307, - "total_runtime_ms": 279, - "weighted_score": 0.9362 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 765, - "total_runtime_ms": 695, - "weighted_score": 0.8295 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.776, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 429, - "total_runtime_ms": 390, - "weighted_score": 0.8411 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1455, - "total_runtime_ms": 1323, - "weighted_score": 0.7401 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.873, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 751, - "total_runtime_ms": 683, - "weighted_score": 0.8272 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2112, - "total_runtime_ms": 1920, - "weighted_score": 0.6743 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.752, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1087, - "total_runtime_ms": 988, - "weighted_score": 0.7197 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:0.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 509, - "total_runtime_ms": 462, - "weighted_score": 0.8629 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:0.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.776, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 267, - "total_runtime_ms": 243, - "weighted_score": 0.8651 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:1.5b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 637, - "total_runtime_ms": 579, - "weighted_score": 0.8833 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:1.5b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.776, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 368, - "total_runtime_ms": 335, - "weighted_score": 0.8505 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:7b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1206, - "total_runtime_ms": 1096, - "weighted_score": 0.8094 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:7b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.873, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 657, - "total_runtime_ms": 598, - "weighted_score": 0.8434 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:14b|system_selected|cold", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.91, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1780, - "total_runtime_ms": 1618, - "weighted_score": 0.7176 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "ambiguous_intent|high|qwen2:14b|system_selected|warm", - "scenario": "ambiguous_intent", - "hardware_profile": "high", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.658, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 928, - "total_runtime_ms": 844, - "weighted_score": 0.708 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:0.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1129, - "total_runtime_ms": 1026, - "weighted_score": 0.8242 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:0.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 689, - "total_runtime_ms": 627, - "weighted_score": 0.8862 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:1.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1507, - "total_runtime_ms": 1370, - "weighted_score": 0.7394 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:1.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.873, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 899, - "total_runtime_ms": 818, - "weighted_score": 0.817 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:7b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2980, - "total_runtime_ms": 2709, - "weighted_score": 0.6414 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:7b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.776, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 1708, - "total_runtime_ms": 1553, - "weighted_score": 0.6611 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:14b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.819, - "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 4378, - "total_runtime_ms": 3979, - "weighted_score": 0.5521 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|low|qwen2:14b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "low", - "model": "qwen2:14b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2526, - "total_runtime_ms": 2296, - "weighted_score": 0.6065 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:0.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.94, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 855, - "total_runtime_ms": 777, - "weighted_score": 0.8584 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:0.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:0.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 525, - "total_runtime_ms": 477, - "weighted_score": 0.9093 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:1.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1149, - "total_runtime_ms": 1045, - "weighted_score": 0.7843 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:1.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:1.5b", - "selection_mode": "system_selected", - "cold_or_warm": "warm", - "runs": 10, - "aggregate": { - "task_success_rate": 0.97, - "schema_validity_rate": 0.98, - "failure_rate": 0.03, - "p95_latency_ms": 702, - "total_runtime_ms": 638, - "weighted_score": 0.8846 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:7b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", - "selection_mode": "system_selected", - "cold_or_warm": "cold", - "runs": 10, - "aggregate": { - "task_success_rate": 0.846, - "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2261, - "total_runtime_ms": 2055, - "weighted_score": 0.6459 - } - }, - { - "evidence_class": "deterministic_simulation", - "runtime_verified": false, - "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:7b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "analyze_reasoning|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.873, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1305, - "total_runtime_ms": 1186, - "weighted_score": 0.7583 + "p95_latency_ms": 397, + "total_runtime_ms": 361, + "weighted_score": 0.9297 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:14b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "analyze_reasoning", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.819, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3307, - "total_runtime_ms": 3007, - "weighted_score": 0.579 + "failure_rate": 0.06, + "p95_latency_ms": 963, + "total_runtime_ms": 876, + "weighted_score": 0.8463 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium|qwen2:14b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "analyze_reasoning|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "analyze_reasoning", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1925, - "total_runtime_ms": 1750, - "weighted_score": 0.6909 + "failure_rate": 0.03, + "p95_latency_ms": 594, + "total_runtime_ms": 540, + "weighted_score": 0.9012 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "compare_structured|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7374,79 +2894,79 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 692, - "total_runtime_ms": 629, - "weighted_score": 0.8797 + "p95_latency_ms": 1228, + "total_runtime_ms": 1117, + "weighted_score": 0.8143 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "compare_structured|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.873, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 408, - "total_runtime_ms": 371, - "weighted_score": 0.8872 + "p95_latency_ms": 757, + "total_runtime_ms": 688, + "weighted_score": 0.8794 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "compare_structured|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.846, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 877, - "total_runtime_ms": 797, - "weighted_score": 0.8183 + "p95_latency_ms": 2010, + "total_runtime_ms": 1828, + "weighted_score": 0.7173 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "compare_structured|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "compare_structured", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.873, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 543, - "total_runtime_ms": 493, - "weighted_score": 0.8685 + "p95_latency_ms": 1205, + "total_runtime_ms": 1096, + "weighted_score": 0.8158 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "compare_structured|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7454,79 +2974,79 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1699, - "total_runtime_ms": 1544, - "weighted_score": 0.7533 + "p95_latency_ms": 922, + "total_runtime_ms": 838, + "weighted_score": 0.8516 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "compare_structured|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.679, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 990, - "total_runtime_ms": 900, - "weighted_score": 0.7257 + "p95_latency_ms": 574, + "total_runtime_ms": 522, + "weighted_score": 0.9044 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "compare_structured|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.819, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2489, - "total_runtime_ms": 2263, - "weighted_score": 0.6002 + "failure_rate": 0.06, + "p95_latency_ms": 1522, + "total_runtime_ms": 1383, + "weighted_score": 0.7776 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "compare_structured|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "compare_structured", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.752, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1437, - "total_runtime_ms": 1307, - "weighted_score": 0.6846 + "failure_rate": 0.03, + "p95_latency_ms": 926, + "total_runtime_ms": 842, + "weighted_score": 0.8551 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:0.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7534,19 +3054,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 572, - "total_runtime_ms": 520, - "weighted_score": 0.8942 + "p95_latency_ms": 699, + "total_runtime_ms": 635, + "weighted_score": 0.879 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:0.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "compare_structured|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -7554,19 +3074,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 378, - "total_runtime_ms": 343, - "weighted_score": 0.9316 + "p95_latency_ms": 451, + "total_runtime_ms": 410, + "weighted_score": 0.9217 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:1.5b|system_selected|cold", - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7574,39 +3094,39 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 731, - "total_runtime_ms": 665, - "weighted_score": 0.8739 + "p95_latency_ms": 1149, + "total_runtime_ms": 1045, + "weighted_score": 0.8234 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:1.5b|system_selected|warm", - "scenario": "contradictory_intent", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "compare_structured|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "compare_structured", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.873, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 453, - "total_runtime_ms": 412, - "weighted_score": 0.8809 + "p95_latency_ms": 711, + "total_runtime_ms": 646, + "weighted_score": 0.8852 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:7b|system_selected|cold", - "scenario": "contradictory_intent", + "combo_id": "compare_structured|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7614,79 +3134,79 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1427, - "total_runtime_ms": 1298, - "weighted_score": 0.7872 + "p95_latency_ms": 565, + "total_runtime_ms": 514, + "weighted_score": 0.8949 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:7b|system_selected|warm", - "scenario": "contradictory_intent", + "combo_id": "compare_structured|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.873, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 854, - "total_runtime_ms": 777, - "weighted_score": 0.8238 + "p95_latency_ms": 400, + "total_runtime_ms": 364, + "weighted_score": 0.9294 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:14b|system_selected|cold", - "scenario": "contradictory_intent", + "combo_id": "compare_structured|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2075, - "total_runtime_ms": 1886, - "weighted_score": 0.6881 + "failure_rate": 0.06, + "p95_latency_ms": 979, + "total_runtime_ms": 890, + "weighted_score": 0.8447 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "contradictory_intent|high|qwen2:14b|system_selected|warm", - "scenario": "contradictory_intent", + "combo_id": "compare_structured|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "compare_structured", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.752, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1230, - "total_runtime_ms": 1118, - "weighted_score": 0.7154 + "failure_rate": 0.03, + "p95_latency_ms": 608, + "total_runtime_ms": 552, + "weighted_score": 0.8998 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:0.5b|system_selected|cold", - "scenario": "long_context", + "combo_id": "ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7694,39 +3214,39 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1342, - "total_runtime_ms": 1220, - "weighted_score": 0.7935 + "p95_latency_ms": 973, + "total_runtime_ms": 885, + "weighted_score": 0.8398 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:0.5b|system_selected|warm", - "scenario": "long_context", + "combo_id": "ambiguous_intent|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.97, + "task_success_rate": 0.873, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 884, - "total_runtime_ms": 804, - "weighted_score": 0.8573 + "p95_latency_ms": 505, + "total_runtime_ms": 459, + "weighted_score": 0.8658 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:1.5b|system_selected|cold", - "scenario": "long_context", + "combo_id": "ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7734,19 +3254,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1808, - "total_runtime_ms": 1643, - "weighted_score": 0.7343 + "p95_latency_ms": 1620, + "total_runtime_ms": 1473, + "weighted_score": 0.7563 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:1.5b|system_selected|warm", - "scenario": "long_context", + "combo_id": "ambiguous_intent|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -7754,39 +3274,39 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1180, - "total_runtime_ms": 1073, - "weighted_score": 0.815 + "p95_latency_ms": 838, + "total_runtime_ms": 762, + "weighted_score": 0.8525 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:7b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.846, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 3519, - "total_runtime_ms": 3199, - "weighted_score": 0.616 + "p95_latency_ms": 764, + "total_runtime_ms": 695, + "weighted_score": 0.8298 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:7b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "ambiguous_intent|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -7794,59 +3314,59 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 2291, - "total_runtime_ms": 2082, - "weighted_score": 0.655 + "p95_latency_ms": 411, + "total_runtime_ms": 374, + "weighted_score": 0.9208 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:14b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.846, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 5199, - "total_runtime_ms": 4727, - "weighted_score": 0.5509 + "failure_rate": 0.06, + "p95_latency_ms": 1226, + "total_runtime_ms": 1115, + "weighted_score": 0.7695 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|low|qwen2:14b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "ambiguous_intent|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.873, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 3381, - "total_runtime_ms": 3074, - "weighted_score": 0.5689 + "failure_rate": 0.03, + "p95_latency_ms": 654, + "total_runtime_ms": 595, + "weighted_score": 0.8435 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:0.5b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7854,39 +3374,39 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1008, - "total_runtime_ms": 917, - "weighted_score": 0.8359 + "p95_latency_ms": 588, + "total_runtime_ms": 535, + "weighted_score": 0.89 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:0.5b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "ambiguous_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.97, + "task_success_rate": 0.679, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 694, - "total_runtime_ms": 631, - "weighted_score": 0.8853 + "p95_latency_ms": 323, + "total_runtime_ms": 294, + "weighted_score": 0.8182 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:1.5b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7894,19 +3414,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1367, - "total_runtime_ms": 1242, - "weighted_score": 0.7906 + "p95_latency_ms": 907, + "total_runtime_ms": 825, + "weighted_score": 0.8475 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:1.5b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "ambiguous_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -7914,19 +3434,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 903, - "total_runtime_ms": 820, - "weighted_score": 0.855 + "p95_latency_ms": 504, + "total_runtime_ms": 458, + "weighted_score": 0.9059 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:7b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -7934,119 +3454,119 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 2662, - "total_runtime_ms": 2420, - "weighted_score": 0.6405 + "p95_latency_ms": 482, + "total_runtime_ms": 439, + "weighted_score": 0.9032 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:7b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "ambiguous_intent|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.97, + "task_success_rate": 0.679, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1743, - "total_runtime_ms": 1585, - "weighted_score": 0.7342 + "p95_latency_ms": 278, + "total_runtime_ms": 253, + "weighted_score": 0.8252 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:14b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 3914, - "total_runtime_ms": 3558, - "weighted_score": 0.5871 + "failure_rate": 0.06, + "p95_latency_ms": 780, + "total_runtime_ms": 709, + "weighted_score": 0.8646 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium|qwen2:14b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "ambiguous_intent|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "ambiguous_intent", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 2542, - "total_runtime_ms": 2311, - "weighted_score": 0.6051 + "failure_rate": 0.03, + "p95_latency_ms": 413, + "total_runtime_ms": 376, + "weighted_score": 0.9193 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:0.5b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "contradictory_intent|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.846, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 772, - "total_runtime_ms": 702, - "weighted_score": 0.8664 + "p95_latency_ms": 1129, + "total_runtime_ms": 1027, + "weighted_score": 0.7866 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:0.5b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "combo_id": "contradictory_intent|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.97, + "task_success_rate": 0.776, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 538, - "total_runtime_ms": 489, - "weighted_score": 0.9078 + "p95_latency_ms": 686, + "total_runtime_ms": 624, + "weighted_score": 0.8089 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:1.5b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "contradictory_intent|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8054,19 +3574,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1038, - "total_runtime_ms": 943, - "weighted_score": 0.8327 + "p95_latency_ms": 1891, + "total_runtime_ms": 1719, + "weighted_score": 0.7293 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:1.5b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "combo_id": "contradictory_intent|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8074,19 +3594,19 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 701, - "total_runtime_ms": 637, - "weighted_score": 0.8844 + "p95_latency_ms": 1111, + "total_runtime_ms": 1010, + "weighted_score": 0.8252 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:7b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8094,79 +3614,79 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 2022, - "total_runtime_ms": 1839, - "weighted_score": 0.7067 + "p95_latency_ms": 860, + "total_runtime_ms": 781, + "weighted_score": 0.8579 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:7b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "combo_id": "contradictory_intent|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.97, + "task_success_rate": 0.776, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1320, - "total_runtime_ms": 1200, - "weighted_score": 0.7949 + "p95_latency_ms": 539, + "total_runtime_ms": 490, + "weighted_score": 0.8303 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:14b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.752, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2954, - "total_runtime_ms": 2685, - "weighted_score": 0.6143 + "failure_rate": 0.06, + "p95_latency_ms": 1433, + "total_runtime_ms": 1303, + "weighted_score": 0.7112 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|medium_gpu|qwen2:14b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "combo_id": "contradictory_intent|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.873, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1922, - "total_runtime_ms": 1747, - "weighted_score": 0.6902 + "failure_rate": 0.03, + "p95_latency_ms": 831, + "total_runtime_ms": 755, + "weighted_score": 0.8259 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:0.5b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8174,39 +3694,39 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 654, - "total_runtime_ms": 595, - "weighted_score": 0.8816 + "p95_latency_ms": 664, + "total_runtime_ms": 603, + "weighted_score": 0.8825 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:0.5b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "combo_id": "contradictory_intent|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.97, + "task_success_rate": 0.873, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 454, - "total_runtime_ms": 413, - "weighted_score": 0.9195 + "p95_latency_ms": 416, + "total_runtime_ms": 378, + "weighted_score": 0.8865 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:1.5b|system_selected|cold", - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8214,19 +3734,19 @@ "task_success_rate": 0.94, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 887, - "total_runtime_ms": 807, - "weighted_score": 0.8523 + "p95_latency_ms": 1069, + "total_runtime_ms": 972, + "weighted_score": 0.8314 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:1.5b|system_selected|warm", - "scenario": "long_context", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "combo_id": "contradictory_intent|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8234,419 +3754,419 @@ "task_success_rate": 0.97, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 574, - "total_runtime_ms": 522, - "weighted_score": 0.9016 + "p95_latency_ms": 673, + "total_runtime_ms": 612, + "weighted_score": 0.889 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:7b|system_selected|cold", - "scenario": "long_context", + "combo_id": "contradictory_intent|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.752, "schema_validity_rate": 0.98, "failure_rate": 0.06, - "p95_latency_ms": 1677, - "total_runtime_ms": 1524, - "weighted_score": 0.7504 + "p95_latency_ms": 581, + "total_runtime_ms": 528, + "weighted_score": 0.8181 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:7b|system_selected|warm", - "scenario": "long_context", + "combo_id": "contradictory_intent|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.97, + "task_success_rate": 0.776, "schema_validity_rate": 0.98, "failure_rate": 0.03, - "p95_latency_ms": 1123, - "total_runtime_ms": 1021, - "weighted_score": 0.8238 + "p95_latency_ms": 378, + "total_runtime_ms": 344, + "weighted_score": 0.854 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:14b|system_selected|cold", - "scenario": "long_context", + "combo_id": "contradictory_intent|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.658, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 2473, - "total_runtime_ms": 2248, - "weighted_score": 0.6307 + "failure_rate": 0.06, + "p95_latency_ms": 898, + "total_runtime_ms": 816, + "weighted_score": 0.74 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "long_context|high|qwen2:14b|system_selected|warm", - "scenario": "long_context", + "combo_id": "contradictory_intent|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "contradictory_intent", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.94, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.06, - "p95_latency_ms": 1618, - "total_runtime_ms": 1471, - "weighted_score": 0.7342 + "failure_rate": 0.03, + "p95_latency_ms": 557, + "total_runtime_ms": 507, + "weighted_score": 0.9049 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:0.5b|system_selected|cold", - "scenario": "noise_input", + "combo_id": "long_context|low|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 960, - "total_runtime_ms": 873, - "weighted_score": 0.8051 + "failure_rate": 0.06, + "p95_latency_ms": 1360, + "total_runtime_ms": 1237, + "weighted_score": 0.7917 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:0.5b|system_selected|warm", - "scenario": "noise_input", + "combo_id": "long_context|low|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:0.5b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 526, - "total_runtime_ms": 478, - "weighted_score": 0.8666 + "failure_rate": 0.03, + "p95_latency_ms": 897, + "total_runtime_ms": 815, + "weighted_score": 0.856 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:1.5b|system_selected|cold", - "scenario": "noise_input", + "combo_id": "long_context|low|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 1294, - "total_runtime_ms": 1176, - "weighted_score": 0.7624 + "failure_rate": 0.06, + "p95_latency_ms": 2258, + "total_runtime_ms": 2053, + "weighted_score": 0.6765 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:1.5b|system_selected|warm", - "scenario": "noise_input", + "combo_id": "long_context|low|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "long_context", "hardware_profile": "low", - "model": "qwen2:1.5b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 690, - "total_runtime_ms": 627, - "weighted_score": 0.8407 + "failure_rate": 0.03, + "p95_latency_ms": 1445, + "total_runtime_ms": 1313, + "weighted_score": 0.7759 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:7b|system_selected|cold", - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "long_context|medium|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 2528, - "total_runtime_ms": 2299, - "weighted_score": 0.6054 + "failure_rate": 0.06, + "p95_latency_ms": 1031, + "total_runtime_ms": 937, + "weighted_score": 0.8337 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:7b|system_selected|warm", - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:7b", + "combo_id": "long_context|medium|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1283, - "total_runtime_ms": 1166, - "weighted_score": 0.7452 + "failure_rate": 0.03, + "p95_latency_ms": 660, + "total_runtime_ms": 600, + "weighted_score": 0.8888 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:14b|system_selected|cold", - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "long_context|medium|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.85, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.15, - "p95_latency_ms": 3727, - "total_runtime_ms": 3388, - "weighted_score": 0.5525 + "failure_rate": 0.06, + "p95_latency_ms": 1698, + "total_runtime_ms": 1544, + "weighted_score": 0.7479 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|low|qwen2:14b|system_selected|warm", - "scenario": "noise_input", - "hardware_profile": "low", - "model": "qwen2:14b", + "combo_id": "long_context|medium|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "long_context", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 1898, - "total_runtime_ms": 1726, - "weighted_score": 0.6307 + "failure_rate": 0.03, + "p95_latency_ms": 1102, + "total_runtime_ms": 1002, + "weighted_score": 0.8255 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:0.5b|system_selected|cold", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 755, - "total_runtime_ms": 687, - "weighted_score": 0.8323 + "failure_rate": 0.06, + "p95_latency_ms": 778, + "total_runtime_ms": 707, + "weighted_score": 0.8658 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:0.5b|system_selected|warm", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:0.5b", + "combo_id": "long_context|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 421, - "total_runtime_ms": 383, - "weighted_score": 0.8837 + "failure_rate": 0.03, + "p95_latency_ms": 525, + "total_runtime_ms": 477, + "weighted_score": 0.9091 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:1.5b|system_selected|cold", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 985, - "total_runtime_ms": 896, - "weighted_score": 0.8023 + "failure_rate": 0.06, + "p95_latency_ms": 1291, + "total_runtime_ms": 1173, + "weighted_score": 0.8002 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:1.5b|system_selected|warm", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:1.5b", + "combo_id": "long_context|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "long_context", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 549, - "total_runtime_ms": 499, - "weighted_score": 0.8639 + "failure_rate": 0.03, + "p95_latency_ms": 869, + "total_runtime_ms": 790, + "weighted_score": 0.8604 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:7b|system_selected|cold", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "long_context|high|gemma4:e2b-it-qat|system_selected|cold", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 1924, - "total_runtime_ms": 1749, - "weighted_score": 0.6812 + "failure_rate": 0.06, + "p95_latency_ms": 663, + "total_runtime_ms": 603, + "weighted_score": 0.8807 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:7b|system_selected|warm", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:7b", + "combo_id": "long_context|high|gemma4:e2b-it-qat|system_selected|warm", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.91, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.09, - "p95_latency_ms": 1010, - "total_runtime_ms": 918, - "weighted_score": 0.7906 + "failure_rate": 0.03, + "p95_latency_ms": 468, + "total_runtime_ms": 426, + "weighted_score": 0.9182 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:14b|system_selected|cold", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "long_context|high|gemma4:e4b-it-qat|system_selected|cold", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.85, + "task_success_rate": 0.94, "schema_validity_rate": 0.98, - "failure_rate": 0.15, - "p95_latency_ms": 2811, - "total_runtime_ms": 2555, - "weighted_score": 0.5794 + "failure_rate": 0.06, + "p95_latency_ms": 1068, + "total_runtime_ms": 971, + "weighted_score": 0.8283 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium|qwen2:14b|system_selected|warm", - "scenario": "noise_input", - "hardware_profile": "medium", - "model": "qwen2:14b", + "combo_id": "long_context|high|gemma4:e4b-it-qat|system_selected|warm", + "scenario": "long_context", + "hardware_profile": "high", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.97, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 1426, - "total_runtime_ms": 1296, - "weighted_score": 0.7048 + "failure_rate": 0.03, + "p95_latency_ms": 715, + "total_runtime_ms": 650, + "weighted_score": 0.8815 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:0.5b|system_selected|cold", + "combo_id": "noise_input|low|gemma4:e2b-it-qat|system_selected|cold", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8654,19 +4174,19 @@ "task_success_rate": 0.88, "schema_validity_rate": 0.98, "failure_rate": 0.12, - "p95_latency_ms": 571, - "total_runtime_ms": 519, - "weighted_score": 0.8558 + "p95_latency_ms": 995, + "total_runtime_ms": 905, + "weighted_score": 0.8016 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:0.5b|system_selected|warm", + "combo_id": "noise_input|low|gemma4:e2b-it-qat|system_selected|warm", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:0.5b", + "hardware_profile": "low", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8674,19 +4194,19 @@ "task_success_rate": 0.91, "schema_validity_rate": 0.98, "failure_rate": 0.09, - "p95_latency_ms": 322, - "total_runtime_ms": 293, - "weighted_score": 0.8987 + "p95_latency_ms": 519, + "total_runtime_ms": 472, + "weighted_score": 0.8672 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:1.5b|system_selected|cold", + "combo_id": "noise_input|low|gemma4:e4b-it-qat|system_selected|cold", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8694,19 +4214,19 @@ "task_success_rate": 0.88, "schema_validity_rate": 0.98, "failure_rate": 0.12, - "p95_latency_ms": 755, - "total_runtime_ms": 686, - "weighted_score": 0.8321 + "p95_latency_ms": 1611, + "total_runtime_ms": 1465, + "weighted_score": 0.7212 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:1.5b|system_selected|warm", + "combo_id": "noise_input|low|gemma4:e4b-it-qat|system_selected|warm", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:1.5b", + "hardware_profile": "low", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8714,19 +4234,19 @@ "task_success_rate": 0.91, "schema_validity_rate": 0.98, "failure_rate": 0.09, - "p95_latency_ms": 411, - "total_runtime_ms": 374, - "weighted_score": 0.8845 + "p95_latency_ms": 848, + "total_runtime_ms": 770, + "weighted_score": 0.8156 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:7b|system_selected|cold", + "combo_id": "noise_input|medium|gemma4:e2b-it-qat|system_selected|cold", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8734,19 +4254,19 @@ "task_success_rate": 0.88, "schema_validity_rate": 0.98, "failure_rate": 0.12, - "p95_latency_ms": 1440, - "total_runtime_ms": 1309, - "weighted_score": 0.7432 + "p95_latency_ms": 780, + "total_runtime_ms": 709, + "weighted_score": 0.8298 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:7b|system_selected|warm", + "combo_id": "noise_input|medium|gemma4:e2b-it-qat|system_selected|warm", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:7b", + "hardware_profile": "medium", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8754,59 +4274,59 @@ "task_success_rate": 0.91, "schema_validity_rate": 0.98, "failure_rate": 0.09, - "p95_latency_ms": 738, - "total_runtime_ms": 671, - "weighted_score": 0.8314 + "p95_latency_ms": 421, + "total_runtime_ms": 382, + "weighted_score": 0.8838 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:14b|system_selected|cold", + "combo_id": "noise_input|medium|gemma4:e4b-it-qat|system_selected|cold", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.85, + "task_success_rate": 0.88, "schema_validity_rate": 0.98, - "failure_rate": 0.15, - "p95_latency_ms": 2117, - "total_runtime_ms": 1925, - "weighted_score": 0.6378 + "failure_rate": 0.12, + "p95_latency_ms": 1241, + "total_runtime_ms": 1128, + "weighted_score": 0.7697 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|medium_gpu|qwen2:14b|system_selected|warm", + "combo_id": "noise_input|medium|gemma4:e4b-it-qat|system_selected|warm", "scenario": "noise_input", - "hardware_profile": "medium_gpu", - "model": "qwen2:14b", + "hardware_profile": "medium", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.91, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 1097, - "total_runtime_ms": 997, - "weighted_score": 0.7578 + "failure_rate": 0.09, + "p95_latency_ms": 659, + "total_runtime_ms": 599, + "weighted_score": 0.8458 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:0.5b|system_selected|cold", + "combo_id": "noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|cold", "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8814,19 +4334,19 @@ "task_success_rate": 0.88, "schema_validity_rate": 0.98, "failure_rate": 0.12, - "p95_latency_ms": 493, - "total_runtime_ms": 448, - "weighted_score": 0.8661 + "p95_latency_ms": 566, + "total_runtime_ms": 514, + "weighted_score": 0.8563 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:0.5b|system_selected|warm", + "combo_id": "noise_input|medium_gpu|gemma4:e2b-it-qat|system_selected|warm", "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:0.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8834,19 +4354,19 @@ "task_success_rate": 0.91, "schema_validity_rate": 0.98, "failure_rate": 0.09, - "p95_latency_ms": 305, - "total_runtime_ms": 277, - "weighted_score": 0.9029 + "p95_latency_ms": 303, + "total_runtime_ms": 275, + "weighted_score": 0.9006 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:1.5b|system_selected|cold", + "combo_id": "noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|cold", "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8854,19 +4374,19 @@ "task_success_rate": 0.88, "schema_validity_rate": 0.98, "failure_rate": 0.12, - "p95_latency_ms": 639, - "total_runtime_ms": 581, - "weighted_score": 0.8471 + "p95_latency_ms": 951, + "total_runtime_ms": 864, + "weighted_score": 0.8072 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:1.5b|system_selected|warm", + "combo_id": "noise_input|medium_gpu|gemma4:e4b-it-qat|system_selected|warm", "scenario": "noise_input", - "hardware_profile": "high", - "model": "qwen2:1.5b", + "hardware_profile": "medium_gpu", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8874,19 +4394,19 @@ "task_success_rate": 0.91, "schema_validity_rate": 0.98, "failure_rate": 0.09, - "p95_latency_ms": 351, - "total_runtime_ms": 319, - "weighted_score": 0.8939 + "p95_latency_ms": 486, + "total_runtime_ms": 442, + "weighted_score": 0.8717 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:7b|system_selected|cold", + "combo_id": "noise_input|high|gemma4:e2b-it-qat|system_selected|cold", "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, @@ -8894,19 +4414,19 @@ "task_success_rate": 0.88, "schema_validity_rate": 0.98, "failure_rate": 0.12, - "p95_latency_ms": 1217, - "total_runtime_ms": 1106, - "weighted_score": 0.7723 + "p95_latency_ms": 508, + "total_runtime_ms": 462, + "weighted_score": 0.8646 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:7b|system_selected|warm", + "combo_id": "noise_input|high|gemma4:e2b-it-qat|system_selected|warm", "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:7b", + "model": "gemma4:e2b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, @@ -8914,49 +4434,49 @@ "task_success_rate": 0.91, "schema_validity_rate": 0.98, "failure_rate": 0.09, - "p95_latency_ms": 652, - "total_runtime_ms": 593, - "weighted_score": 0.8468 + "p95_latency_ms": 267, + "total_runtime_ms": 243, + "weighted_score": 0.9067 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:14b|system_selected|cold", + "combo_id": "noise_input|high|gemma4:e4b-it-qat|system_selected|cold", "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "cold", "runs": 10, "aggregate": { - "task_success_rate": 0.85, + "task_success_rate": 0.88, "schema_validity_rate": 0.98, - "failure_rate": 0.15, - "p95_latency_ms": 1783, - "total_runtime_ms": 1621, - "weighted_score": 0.6813 + "failure_rate": 0.12, + "p95_latency_ms": 798, + "total_runtime_ms": 725, + "weighted_score": 0.8268 } }, { "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "combo_id": "noise_input|high|qwen2:14b|system_selected|warm", + "combo_id": "noise_input|high|gemma4:e4b-it-qat|system_selected|warm", "scenario": "noise_input", "hardware_profile": "high", - "model": "qwen2:14b", + "model": "gemma4:e4b-it-qat", "selection_mode": "system_selected", "cold_or_warm": "warm", "runs": 10, "aggregate": { - "task_success_rate": 0.88, + "task_success_rate": 0.91, "schema_validity_rate": 0.98, - "failure_rate": 0.12, - "p95_latency_ms": 941, - "total_runtime_ms": 855, - "weighted_score": 0.7835 + "failure_rate": 0.09, + "p95_latency_ms": 410, + "total_runtime_ms": 372, + "weighted_score": 0.8836 } } ] diff --git a/reports/simulation_report.json b/reports/simulation_report.json index 4d71049..bfa9465 100644 --- a/reports/simulation_report.json +++ b/reports/simulation_report.json @@ -2,7 +2,7 @@ "evidence_class": "deterministic_simulation", "runtime_verified": false, "promotion_eligible": false, - "generated_at": "2026-08-18T19:23:33.758Z", + "generated_at": "2026-08-25T17:30:23.118Z", "scenario_matrix": { "dimensions": [ "task_fixture", diff --git a/reports/spec_coverage_report.json b/reports/spec_coverage_report.json index fa74412..42c2511 100644 --- a/reports/spec_coverage_report.json +++ b/reports/spec_coverage_report.json @@ -1,5 +1,5 @@ { - "generated_at": "2026-08-18T19:23:33.758Z", + "generated_at": "2026-08-25T17:30:23.118Z", "implemented_spec_nodes": [ "dynamic_model_evaluation_layer", "execution_layer", diff --git a/reports/verification_report.json b/reports/verification_report.json index 06d1f5d..03ebb92 100644 --- a/reports/verification_report.json +++ b/reports/verification_report.json @@ -1,5 +1,5 @@ { - "generated_at": "2026-08-18T19:23:33.758Z", + "generated_at": "2026-08-25T17:30:23.118Z", "verification_scope": "static_traceability_only", "full_system_ok": false, "runtime_verified": false, diff --git a/runtime/model_policy.json b/runtime/model_policy.json index 2dd9426..3b2babd 100644 --- a/runtime/model_policy.json +++ b/runtime/model_policy.json @@ -1,6 +1,6 @@ { - "policy_version": "runtime-policy-c44de2e4a8a1", - "generated_at_unix_ms": 1700798162081, + "policy_version": "runtime-policy-8226a11e6ae3", + "generated_at_unix_ms": 1700423493347, "promotion_evidence": { "runtime_verified": false, "status": "simulation_only_no_promotion" @@ -22,172 +22,145 @@ "task_family": "extract", "hardware_profile": "low", "output_mode": "strict_json", - "preferred_model_id": "qwen2.5:0.5b", + "preferred_model_id": "gemma4:e2b-it-qat", "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "extract", "hardware_profile": "medium", "output_mode": "strict_json", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b" - ], + "preferred_model_id": "gemma4:e2b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "extract", "hardware_profile": "medium_gpu", "output_mode": "strict_json", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b", - "qwen2.5:7b" - ], + "preferred_model_id": "gemma4:e2b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "extract", "hardware_profile": "high", "output_mode": "strict_json", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b", - "qwen2.5:7b", - "qwen2.5:14b" - ], + "preferred_model_id": "gemma4:e2b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "summarize", "hardware_profile": "low", "output_mode": "semi_structured", - "preferred_model_id": "qwen2.5:0.5b", + "preferred_model_id": "gemma4:e2b-it-qat", "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "summarize", "hardware_profile": "medium", "output_mode": "semi_structured", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b" - ], + "preferred_model_id": "gemma4:e2b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "summarize", "hardware_profile": "medium_gpu", "output_mode": "semi_structured", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b", - "qwen2.5:7b" - ], + "preferred_model_id": "gemma4:e2b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "summarize", "hardware_profile": "high", "output_mode": "semi_structured", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b", - "qwen2.5:7b", - "qwen2.5:14b" - ], + "preferred_model_id": "gemma4:e2b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "agent", "hardware_profile": "low", "output_mode": "freeform", - "preferred_model_id": "qwen2.5:0.5b", + "preferred_model_id": "gemma4:e2b-it-qat", "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "agent", "hardware_profile": "medium", "output_mode": "freeform", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b" - ], + "preferred_model_id": "gemma4:e4b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "agent", "hardware_profile": "medium_gpu", "output_mode": "freeform", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b", - "qwen2.5:7b" - ], + "preferred_model_id": "gemma4:e4b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 }, { "task_family": "agent", "hardware_profile": "high", "output_mode": "freeform", - "preferred_model_id": "qwen2.5:0.5b", - "fallback_model_ids": [ - "qwen2.5:1.5b", - "qwen2.5:7b", - "qwen2.5:14b" - ], + "preferred_model_id": "gemma4:e4b-it-qat", + "fallback_model_ids": [], "selection_reason": "baseline_registry_no_runtime_promotion", "evidence_refs": [ "server/model/registry_source.json" ], - "effective_from_unix_ms": 1700798162081 + "effective_from_unix_ms": 1700423493347 } ], "quarantined_models": [], diff --git a/runtime/model_registry.runtime.json b/runtime/model_registry.runtime.json index 15d8e40..29ef70b 100644 --- a/runtime/model_registry.runtime.json +++ b/runtime/model_registry.runtime.json @@ -1,70 +1,61 @@ { - "generated_at_unix_ms": 1700798162081, - "policy_version": "runtime-policy-c44de2e4a8a1", + "generated_at_unix_ms": 1700423493347, + "policy_version": "runtime-policy-8226a11e6ae3", "models": [ { - "model_id": "qwen2.5:0.5b", - "ollama_name": "qwen2.5:0.5b", + "model_id": "gemma4:e2b-it-qat", + "ollama_name": "gemma4:e2b-it-qat", "supported_operation_families": [ "extract", "summarize", "agent" ], + "preferred_task_families": [ + "extract", + "summarize" + ], "min_hardware_profile": "low", "installation_state": "installed", + "baseline_approved": true, "runtime_status": "preferred", - "reliability_score": 0.81, + "reliability_score": 0.9, "policy_refs": [ - "runtime/model_policy.json#qwen2.5:0.5b" + "runtime/model_policy.json#gemma4:e2b-it-qat" ] }, { - "model_id": "qwen2.5:1.5b", - "ollama_name": "qwen2.5:1.5b", + "model_id": "gemma4:e4b-it-qat", + "ollama_name": "gemma4:e4b-it-qat", "supported_operation_families": [ - "extract", - "summarize", + "agent" + ], + "preferred_task_families": [ "agent" ], "min_hardware_profile": "medium", "installation_state": "installed", - "runtime_status": "fallback", - "reliability_score": 0.88, + "baseline_approved": true, + "runtime_status": "preferred", + "reliability_score": 0.95, "policy_refs": [ - "runtime/model_policy.json#qwen2.5:1.5b" + "runtime/model_policy.json#gemma4:e4b-it-qat" ] }, { "model_id": "qwen2.5:7b", "ollama_name": "qwen2.5:7b", "supported_operation_families": [ - "extract", - "summarize", "agent" ], + "preferred_task_families": [], "min_hardware_profile": "medium_gpu", "installation_state": "available", - "runtime_status": "fallback", + "baseline_approved": false, + "runtime_status": "baseline", "reliability_score": 0.91, "policy_refs": [ "runtime/model_policy.json#qwen2.5:7b" ] - }, - { - "model_id": "qwen2.5:14b", - "ollama_name": "qwen2.5:14b", - "supported_operation_families": [ - "extract", - "summarize", - "agent" - ], - "min_hardware_profile": "high", - "installation_state": "available", - "runtime_status": "fallback", - "reliability_score": 0.93, - "policy_refs": [ - "runtime/model_policy.json#qwen2.5:14b" - ] } ] } diff --git a/runtime/promotion_audit.json b/runtime/promotion_audit.json index 00ec460..9e6c9b4 100644 --- a/runtime/promotion_audit.json +++ b/runtime/promotion_audit.json @@ -1,6 +1,6 @@ { - "generated_at_unix_ms": 1700798162081, - "policy_version": "runtime-policy-c44de2e4a8a1", + "generated_at_unix_ms": 1700423493347, + "policy_version": "runtime-policy-8226a11e6ae3", "source_reports": [ "reports/aggregated_metrics.json", "reports/frontier_decisions.json", @@ -8,13 +8,6 @@ "reports/architecture_decision_report.json" ], "events": [ - { - "event_type": "compile_policy", - "generated_at_unix_ms": 1700704147799, - "promoted_count": 1, - "rejected_count": 0, - "policy_version": "runtime-policy-d7833f631d57" - }, { "event_type": "rollback", "generated_at_unix_ms": 1774549560013, @@ -24,13 +17,6 @@ "task_family": "agent", "hardware_profile": "high" }, - { - "event_type": "compile_policy", - "generated_at_unix_ms": 1700704147799, - "promoted_count": 1, - "rejected_count": 0, - "policy_version": "runtime-policy-d7833f631d57" - }, { "event_type": "rollback", "generated_at_unix_ms": 1774549605424, @@ -42,50 +28,11 @@ }, { "event_type": "compile_policy", - "generated_at_unix_ms": 1700080928761, - "promoted_count": 0, - "rejected_count": 0, - "policy_version": "runtime-policy-20fa39a37df9" - }, - { - "event_type": "compile_policy", - "generated_at_unix_ms": 1700080928761, - "promoted_count": 0, - "rejected_count": 0, - "runtime_evidence_verified": false, - "policy_version": "runtime-policy-20fa39a37df9" - }, - { - "event_type": "compile_policy", - "generated_at_unix_ms": 1700890687120, - "promoted_count": 0, - "rejected_count": 0, - "runtime_evidence_verified": false, - "policy_version": "runtime-policy-237557c53290" - }, - { - "event_type": "compile_policy", - "generated_at_unix_ms": 1700611533398, - "promoted_count": 0, - "rejected_count": 0, - "runtime_evidence_verified": false, - "policy_version": "runtime-policy-83fb9814da56" - }, - { - "event_type": "compile_policy", - "generated_at_unix_ms": 1700559313270, - "promoted_count": 0, - "rejected_count": 0, - "runtime_evidence_verified": false, - "policy_version": "runtime-policy-bbc138feed76" - }, - { - "event_type": "compile_policy", - "generated_at_unix_ms": 1700798162081, + "generated_at_unix_ms": 1700423493347, "promoted_count": 0, "rejected_count": 0, "runtime_evidence_verified": false, - "policy_version": "runtime-policy-c44de2e4a8a1" + "policy_version": "runtime-policy-8226a11e6ae3" } ], "gating_results": [] diff --git a/scripts/bootstrap-macos-local.sh b/scripts/bootstrap-macos-local.sh index 7882e39..70657a3 100755 --- a/scripts/bootstrap-macos-local.sh +++ b/scripts/bootstrap-macos-local.sh @@ -63,8 +63,12 @@ print(json.dumps({ "label": runtime_profile.label, "reason": reason, "generation_model": runtime_profile.generation_model, + "fast_generation_model": runtime_profile.fast_generation_model, "embedding_model": runtime_profile.embedding_model, "num_ctx": runtime_profile.num_ctx, + "fast_num_ctx": runtime_profile.fast_num_ctx, + "max_input_chars": runtime_profile.max_input_chars, + "generation_routes": commands["generation_routes"], "command": commands["command"], })) PY @@ -77,12 +81,22 @@ import shlex import sys payload = json.loads(sys.argv[1]) -for key in ("selected_profile", "generation_model", "embedding_model", "num_ctx", "reason"): +for key in ( + "selected_profile", "generation_model", "fast_generation_model", "embedding_model", + "num_ctx", "fast_num_ctx", "max_input_chars", "reason", +): print(f"{key.upper()}={shlex.quote(str(payload[key]))}") PY )" eval "$PROFILE_VARS" +: "${GENERATION_MODEL:?runtime profile did not provide generation_model}" +: "${FAST_GENERATION_MODEL:?runtime profile did not provide fast_generation_model}" +: "${EMBEDDING_MODEL:?runtime profile did not provide embedding_model}" +: "${NUM_CTX:?runtime profile did not provide num_ctx}" +: "${FAST_NUM_CTX:?runtime profile did not provide fast_num_ctx}" +: "${MAX_INPUT_CHARS:?runtime profile did not provide max_input_chars}" + if [[ "$PLAN_ONLY" == "1" ]]; then printf '%s\n' "$PROFILE_JSON" exit 0 @@ -130,8 +144,14 @@ if [[ "$STATUS_OLLAMA_RUNNING" != "ok" ]]; then fi if [[ "$SKIP_MODEL_PULL" != "1" ]]; then - echo "Pulling generation model: $GENERATION_MODEL" - ollama pull "$GENERATION_MODEL" + GENERATION_MODELS=("$FAST_GENERATION_MODEL") + if [[ "$GENERATION_MODEL" != "$FAST_GENERATION_MODEL" ]]; then + GENERATION_MODELS+=("$GENERATION_MODEL") + fi + for model in "${GENERATION_MODELS[@]}"; do + echo "Pulling generation model: $model" + ollama pull "$model" + done echo "Pulling embedding model: $EMBEDDING_MODEL" ollama pull "$EMBEDDING_MODEL" STATUS_MODEL_PULL="ok" @@ -139,41 +159,48 @@ else STATUS_MODEL_PULL="skipped" fi -echo "Preparing generation model: $GENERATION_MODEL" -python3 - "$GENERATION_MODEL" "$NUM_CTX" <<'PY' +echo "Preparing local model bundle" +python3 - "$GENERATION_MODEL" "$NUM_CTX" "$FAST_GENERATION_MODEL" "$FAST_NUM_CTX" <<'PY' import json import sys from urllib.request import Request, urlopen -model = sys.argv[1] -num_ctx = int(sys.argv[2]) -request = Request( - "http://127.0.0.1:11434/api/generate", - data=json.dumps({ - "model": model, - "prompt": "Return only: ready", - "stream": False, - "keep_alive": "10m", - "options": { - "temperature": 0, - "seed": 42, - "num_ctx": num_ctx, - "num_predict": 8, - }, - }).encode("utf-8"), - headers={"Content-Type": "application/json"}, - method="POST", -) -with urlopen(request, timeout=600) as response: - result = json.load(response) -if result.get("done") is not True: - raise SystemExit("Generation model did not finish preparing.") +routes = [(sys.argv[1], int(sys.argv[2])), (sys.argv[3], int(sys.argv[4]))] +seen = set() +for model, num_ctx in reversed(routes): + if model in seen: + continue + seen.add(model) + request = Request( + "http://127.0.0.1:11434/api/generate", + data=json.dumps({ + "model": model, + "prompt": "Return only: ready", + "stream": False, + "keep_alive": "10m", + "options": { + "temperature": 0, + "seed": 42, + "num_ctx": num_ctx, + "num_predict": 8, + }, + }).encode("utf-8"), + headers={"Content-Type": "application/json"}, + method="POST", + ) + with urlopen(request, timeout=600) as response: + result = json.load(response) + if result.get("done") is not True: + raise SystemExit("Generation model did not finish preparing.") PY STATUS_MODEL_WARMUP="ok" CHROMEAI_OLLAMA_MODEL="$GENERATION_MODEL" \ +CHROMEAI_OLLAMA_FAST_MODEL="$FAST_GENERATION_MODEL" \ CHROMEAI_OLLAMA_EMBED_MODEL="$EMBEDDING_MODEL" \ CHROMEAI_OLLAMA_NUM_CTX="$NUM_CTX" \ +CHROMEAI_OLLAMA_FAST_NUM_CTX="$FAST_NUM_CTX" \ +CHROMEAI_MAX_INPUT_CHARS="$MAX_INPUT_CHARS" \ CHROMEAI_OLLAMA_SEED=42 \ "$ROOT/scripts/install-macos-local.sh" STATUS_LAUNCHAGENT="ok" @@ -212,6 +239,7 @@ SelectPilot bootstrap complete. Profile: $SELECTED_PROFILE Reason: $REASON Generation model: $GENERATION_MODEL +Structured model: $FAST_GENERATION_MODEL Embedding model: $EMBEDDING_MODEL Context window: $NUM_CTX diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index cd31309..36fe2c3 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -7,8 +7,11 @@ DEST="${HOME}/Library/LaunchAgents/com.chromeai.nano.plist" HASH="$(shasum -a 256 "$ROOT/server/nano_server.py" | awk '{print $1}')" OLLAMA_BASE_URL="${CHROMEAI_OLLAMA_BASE_URL:-http://127.0.0.1:11434}" OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-gemma4:e2b-it-qat}" +OLLAMA_FAST_MODEL="${CHROMEAI_OLLAMA_FAST_MODEL:-$OLLAMA_MODEL}" OLLAMA_EMBED_MODEL="${CHROMEAI_OLLAMA_EMBED_MODEL:-nomic-embed-text-v2-moe:latest}" OLLAMA_NUM_CTX="${CHROMEAI_OLLAMA_NUM_CTX:-16384}" +OLLAMA_FAST_NUM_CTX="${CHROMEAI_OLLAMA_FAST_NUM_CTX:-$OLLAMA_NUM_CTX}" +MAX_INPUT_CHARS="${CHROMEAI_MAX_INPUT_CHARS:-16000}" OLLAMA_SEED="${CHROMEAI_OLLAMA_SEED:-42}" RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run}" LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}" @@ -23,8 +26,11 @@ sed \ -e "s|__BINARY_HASH__|$HASH|g" \ -e "s|__OLLAMA_BASE_URL__|$OLLAMA_BASE_URL|g" \ -e "s|__OLLAMA_MODEL__|$OLLAMA_MODEL|g" \ + -e "s|__OLLAMA_FAST_MODEL__|$OLLAMA_FAST_MODEL|g" \ -e "s|__OLLAMA_EMBED_MODEL__|$OLLAMA_EMBED_MODEL|g" \ -e "s|__OLLAMA_NUM_CTX__|$OLLAMA_NUM_CTX|g" \ + -e "s|__OLLAMA_FAST_NUM_CTX__|$OLLAMA_FAST_NUM_CTX|g" \ + -e "s|__MAX_INPUT_CHARS__|$MAX_INPUT_CHARS|g" \ -e "s|__OLLAMA_SEED__|$OLLAMA_SEED|g" \ "$TEMPLATE" > "$DEST" @@ -40,10 +46,11 @@ Next steps: $ROOT 2. Current Ollama base URL: $OLLAMA_BASE_URL 3. Current Ollama model: $OLLAMA_MODEL - 4. Local bridge URL: http://127.0.0.1:8083 - 5. Ollama context window: $OLLAMA_NUM_CTX - 6. Deterministic seed: $OLLAMA_SEED - 7. Run dir: $RUN_DIR - 8. Log dir: $LOG_DIR - 9. Run 'pnpm benchmark:local' to validate latency on this machine. + 4. Structured-task model: $OLLAMA_FAST_MODEL + 5. Local bridge URL: http://127.0.0.1:8083 + 6. Ollama context window: $OLLAMA_NUM_CTX + 7. Deterministic seed: $OLLAMA_SEED + 8. Run dir: $RUN_DIR + 9. Log dir: $LOG_DIR + 10. Run 'pnpm benchmark:local' to validate latency on this machine. EOF diff --git a/scripts/runtime-policy-lib.mjs b/scripts/runtime-policy-lib.mjs index e1b0633..ac584a2 100644 --- a/scripts/runtime-policy-lib.mjs +++ b/scripts/runtime-policy-lib.mjs @@ -250,8 +250,10 @@ export function buildRuntimeRegistryFromPolicy(policy, registrySource, rejectedC model_id: model.model_id, ollama_name: model.ollama_name, supported_operation_families: model.supported_operation_families || ['extract', 'summarize', 'agent'], + preferred_task_families: model.preferred_task_families || [], min_hardware_profile: model.min_hardware_profile || 'low', installation_state: model.installation_state || 'installed', + baseline_approved: model.baseline_approved === true, runtime_status: runtimeStatus, reliability_score: model.reliability_score ?? null, policy_refs: [`runtime/model_policy.json#${model.model_id}`], @@ -266,13 +268,19 @@ function pickOutputMode(taskFamily) { return 'freeform'; } -function modelTaskSupport(modelId) { - return { - extract: true, - summarize: true, - agent: true, - [modelId]: true, - }; +function modelTaskSupport(model) { + const supported = Array.isArray(model.supported_operation_families) + ? model.supported_operation_families + : ['extract', 'summarize', 'agent']; + return new Set(supported); +} + +// @spec_ref prompt_determinism +export function replaceCompileAuditEvent(existingEvents, compileEvent) { + return [ + ...(existingEvents || []).filter((event) => event.event_type !== 'compile_policy'), + compileEvent, + ]; } function confidenceFromDecision(decision, determinismRate) { @@ -427,12 +435,18 @@ export async function compileRuntimePolicy() { const runtimeCandidates = (registrySource.models || []) .filter((m) => !m.quarantined) + .filter((m) => m.baseline_approved === true) .filter((m) => compatibleWithHardware(m.min_hardware_profile || 'low', hardwareProfile)) - .filter((m) => modelTaskSupport(m.model_id)[taskFamily]) + .filter((m) => modelTaskSupport(m).has(taskFamily)) + .sort((a, b) => { + const aPreferred = (a.preferred_task_families || []).includes(taskFamily) ? 0 : 1; + const bPreferred = (b.preferred_task_families || []).includes(taskFamily) ? 0 : 1; + return aPreferred - bPreferred; + }) .map((m) => m.model_id); const preferredInCandidates = preferred && runtimeCandidates.includes(preferred) ? preferred : runtimeCandidates[0] || null; - const fallbackIds = runtimeCandidates.filter((id) => id !== preferredInCandidates).slice(0, 3); + const fallbackIds = []; if (preferredInCandidates) { defaults.push({ @@ -504,17 +518,14 @@ export async function compileRuntimePolicy() { generated_at_unix_ms: generatedAt, policy_version: policy.policy_version, source_reports: policy.source_reports, - events: [ - ...(existingAudit.events || []), - { + events: replaceCompileAuditEvent(existingAudit.events, { event_type: 'compile_policy', generated_at_unix_ms: generatedAt, promoted_count: eligible.length, rejected_count: gatingResults.filter((r) => !r.eligible).length, runtime_evidence_verified: runtimeEvidenceVerified, policy_version: policy.policy_version, - }, - ], + }), gating_results: gatingResults.map((r) => ({ candidate_model: r.candidate_model, eligible: r.eligible, diff --git a/scripts/spec-compiler.mjs b/scripts/spec-compiler.mjs index 4694286..863c352 100644 --- a/scripts/spec-compiler.mjs +++ b/scripts/spec-compiler.mjs @@ -35,7 +35,7 @@ const REQUIRED_TOP_LEVEL_NODES = [ ]; const SCAN_EXTENSIONS = new Set(['.ts', '.tsx', '.js', '.mjs', '.py']); -const SCAN_IGNORE = new Set(['.git', 'node_modules', 'test-results', 'dist']); +const SCAN_IGNORE = new Set(['.git', 'node_modules', 'test-results', 'dist', 'site-dist']); const SIMULATION_EVIDENCE = { evidence_class: 'deterministic_simulation', runtime_verified: false, @@ -102,6 +102,8 @@ function percentile(values, p) { } function modelSizeFactor(model) { + if (model.includes('e2b')) return 1.0; + if (model.includes('e4b')) return 1.7; if (model.includes('0.5b')) return 1.0; if (model.includes('1.5b')) return 1.35; if (model.includes('7b')) return 2.7; diff --git a/selectpilot_benchmark_v1.json b/selectpilot_benchmark_v1.json index a0c5df6..639ab40 100644 --- a/selectpilot_benchmark_v1.json +++ b/selectpilot_benchmark_v1.json @@ -154,10 +154,8 @@ }, "model_matrix": { "candidates": [ - "qwen2:0.5b", - "qwen2:1.5b", - "qwen2:7b", - "qwen2:14b" + "gemma4:e2b-it-qat", + "gemma4:e4b-it-qat" ], "states": [ "cold", @@ -328,4 +326,4 @@ "must remain deterministic" ] } -} \ No newline at end of file +} diff --git a/server/installation_manager.py b/server/installation_manager.py index f347a46..9688f0a 100644 --- a/server/installation_manager.py +++ b/server/installation_manager.py @@ -12,7 +12,7 @@ from pathlib import Path from urllib.request import Request, urlopen -from runtime_profiles import get_runtime_profile, recommend_runtime_profile +from runtime_profiles import get_runtime_profile, recommend_runtime_profile, required_generation_models OLLAMA_DOWNLOAD_URL = "https://ollama.com/download/Ollama-darwin.zip" OLLAMA_TEAM_ID = "3MU9H2V9Y9" @@ -156,16 +156,22 @@ def _run(self) -> None: try: recommendation = recommend_runtime_profile() profile = get_runtime_profile(recommendation["recommended_profile"]) - self._update(profile=profile.key) + generation_models = required_generation_models(profile) + self._update(profile=profile.key, model_bundle=[model for model, _ in generation_models]) ollama = self._install_ollama() self._update(label="Starting local processing", progress=42) subprocess.run(["/usr/bin/open", str(ollama.parents[2]), "--args", "hidden"], check=True) self._wait_for_ollama() self._update(label="Optimizing for this Mac", progress=48) - self._pull_model(profile.generation_model, 48, 82) - self._pull_model(profile.embedding_model, 82, 94) + model_span = 40 / max(1, len(generation_models)) + for index, (model, _) in enumerate(generation_models): + start = round(48 + index * model_span) + end = round(48 + (index + 1) * model_span) + self._pull_model(model, start, end) + self._pull_model(profile.embedding_model, 88, 94) self._update(label="Final checks", progress=96) - self._warm_model(profile.generation_model, profile.num_ctx) + for model, num_ctx in reversed(generation_models): + self._warm_model(model, num_ctx) time.sleep(0.2) self._update(state="ready", label="Go", progress=100) except Exception as error: diff --git a/server/model/registry_source.json b/server/model/registry_source.json index e1c09e4..92a35c4 100644 --- a/server/model/registry_source.json +++ b/server/model/registry_source.json @@ -2,44 +2,37 @@ "version": "runtime-registry-source.v1", "models": [ { - "model_id": "qwen2.5:0.5b", - "ollama_name": "qwen2.5:0.5b", + "model_id": "gemma4:e2b-it-qat", + "ollama_name": "gemma4:e2b-it-qat", "supported_operation_families": ["extract", "summarize", "agent"], + "preferred_task_families": ["extract", "summarize"], "min_hardware_profile": "low", "installation_state": "installed", - "reliability_score": 0.81, + "reliability_score": 0.9, "quarantined": false, "baseline_approved": true }, { - "model_id": "qwen2.5:1.5b", - "ollama_name": "qwen2.5:1.5b", - "supported_operation_families": ["extract", "summarize", "agent"], + "model_id": "gemma4:e4b-it-qat", + "ollama_name": "gemma4:e4b-it-qat", + "supported_operation_families": ["agent"], + "preferred_task_families": ["agent"], "min_hardware_profile": "medium", "installation_state": "installed", - "reliability_score": 0.88, + "reliability_score": 0.95, "quarantined": false, "baseline_approved": true }, { "model_id": "qwen2.5:7b", "ollama_name": "qwen2.5:7b", - "supported_operation_families": ["extract", "summarize", "agent"], + "supported_operation_families": ["agent"], + "preferred_task_families": [], "min_hardware_profile": "medium_gpu", "installation_state": "available", "reliability_score": 0.91, "quarantined": false, - "baseline_approved": true - }, - { - "model_id": "qwen2.5:14b", - "ollama_name": "qwen2.5:14b", - "supported_operation_families": ["extract", "summarize", "agent"], - "min_hardware_profile": "high", - "installation_state": "available", - "reliability_score": 0.93, - "quarantined": false, - "baseline_approved": true + "baseline_approved": false } ] } diff --git a/server/nano_server.py b/server/nano_server.py index 0fa7a13..5d5a954 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -194,9 +194,21 @@ def _expect_optional_dict(payload: dict, field: str) -> dict | None: return raw +def _enforce_local_input_budget(text: str) -> None: + limit = OLLAMA.config.max_input_chars + if len(text) > limit: + raise ValidationError( + "selection_too_large", + "This selection is too large for a predictable local response. Select a smaller section and try again.", + status=413, + details={"maximum_characters": limit, "received_characters": len(text)}, + ) + + def validate_summarize_payload(payload: object) -> dict: body = _expect_dict(payload) - _expect_string(body, "text", required=True, allow_empty=False) + text = _expect_string(body, "text", required=True, allow_empty=False) or "" + _enforce_local_input_budget(text) _expect_string(body, "title", required=False, allow_empty=True) _expect_string(body, "url", required=False, allow_empty=True) _expect_optional_dict(body, "metadata") @@ -205,7 +217,8 @@ def validate_summarize_payload(payload: object) -> dict: def validate_extract_payload(payload: object) -> dict: body = _expect_dict(payload) - _expect_string(body, "text", required=True, allow_empty=False) + text = _expect_string(body, "text", required=True, allow_empty=False) or "" + _enforce_local_input_budget(text) preset = _expect_string(body, "preset", required=False, allow_empty=True) try: get_extraction_preset(preset or None) @@ -224,8 +237,10 @@ def validate_extract_payload(payload: object) -> dict: def validate_agent_payload(payload: object) -> dict: body = _expect_dict(payload) - _expect_string(body, "prompt", required=True, allow_empty=False) - _expect_optional_dict(body, "context") + prompt = _expect_string(body, "prompt", required=True, allow_empty=False) or "" + context = _expect_optional_dict(body, "context") or {} + selected_text = str(context.get("selection") or context.get("pageText") or context.get("markdown") or "") + _enforce_local_input_budget(f"{prompt}\n{selected_text}".strip()) return body @@ -279,6 +294,38 @@ def _validate_object_output(value: object, field: str) -> None: ) +def _validate_routing_output(body: dict) -> None: + routing = body.get("routing") + if not isinstance(routing, dict): + raise ValidationError( + "invalid_model_output", + "routing must be an object", + status=502, + details={"field": "routing", "expected": "object"}, + ) + if not isinstance(routing.get("model"), str) or routing.get("model") != body.get("model"): + raise ValidationError( + "invalid_model_output", + "routing model must match the executed model", + status=502, + details={"field": "routing.model", "expected": body.get("model")}, + ) + if not isinstance(routing.get("reason"), str) or not routing.get("reason"): + raise ValidationError( + "invalid_model_output", + "routing reason must be a non-empty string", + status=502, + details={"field": "routing.reason", "expected": "non-empty string"}, + ) + if not isinstance(routing.get("num_ctx"), int) or routing.get("num_ctx") <= 0: + raise ValidationError( + "invalid_model_output", + "routing context must be a positive integer", + status=502, + details={"field": "routing.num_ctx", "expected": "positive integer"}, + ) + + def validate_summarize_response(response: object) -> dict: body = _expect_dict(response, field="response") for key in ["summary", "markdown", "title", "model", "source", "raw_response"]: @@ -292,6 +339,7 @@ def validate_summarize_response(response: object) -> dict: _validate_list_of_strings(body.get("bullets"), "bullets") _validate_list_of_strings(body.get("action_items"), "action_items") _validate_list_of_strings(body.get("tags"), "tags") + _validate_routing_output(body) return body @@ -307,6 +355,7 @@ def validate_agent_response(response: object) -> dict: ) _validate_list_of_strings(body.get("reasoning"), "reasoning") _validate_object_output(body.get("json"), "json") + _validate_routing_output(body) return body @@ -321,6 +370,7 @@ def validate_extract_response(response: object) -> dict: details={"field": key, "expected": "string"}, ) _validate_object_output(body.get("json"), "json") + _validate_routing_output(body) return body @@ -422,17 +472,17 @@ def write_port_info(port_file: Path, port: int): DETERMINISTIC_MODEL_REGISTRY = [ { - "id": "qwen2.5:0.5b", - "capability_profile": {"classification": 0.9, "extract": 0.82, "rewrite": 0.75, "analyze": 0.72}, + "id": "gemma4:e2b-it-qat", + "capability_profile": {"classification": 0.92, "extract": 0.94, "rewrite": 0.91, "analyze": 0.82}, "resource_profile": {"memory": 1, "latency": 1}, - "benchmark_scores": {"precision": 0.81, "validation_pass_rate": 0.93, "retry_rate": 0.07}, + "benchmark_scores": {"precision": 0.9, "validation_pass_rate": 1.0, "retry_rate": 0.0}, "installation_state": "installed", }, { - "id": "qwen2.5:1.5b", - "capability_profile": {"classification": 0.94, "extract": 0.9, "rewrite": 0.84, "analyze": 0.83}, + "id": "gemma4:e4b-it-qat", + "capability_profile": {"classification": 0.94, "extract": 0.9, "rewrite": 0.88, "analyze": 0.96}, "resource_profile": {"memory": 2, "latency": 2}, - "benchmark_scores": {"precision": 0.88, "validation_pass_rate": 0.95, "retry_rate": 0.05}, + "benchmark_scores": {"precision": 0.95, "validation_pass_rate": 1.0, "retry_rate": 0.0}, "installation_state": "installed", }, ] @@ -783,10 +833,16 @@ def ensure_runtime_models() -> dict: can run with a hardware-fit local model set. """ health = OLLAMA.health() - generation_model = str(health.get("requested_model") or "").strip() + generation_models = [ + str(route.get("model") or "").strip() + for route in (health.get("task_routes") or {}).values() + if isinstance(route, dict) + ] + if not generation_models: + generation_models = [str(health.get("requested_model") or "").strip()] embedding_model = str(health.get("requested_embed_model") or "").strip() - models_to_ensure = [name for name in [generation_model, embedding_model] if name] + models_to_ensure = list(dict.fromkeys(name for name in [*generation_models, embedding_model] if name)) if not models_to_ensure: return {"ok": True, "already_present": [], "pulled": [], "failed": {}} diff --git a/server/ollama_client.py b/server/ollama_client.py index 0abf54b..de7578b 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -101,6 +101,9 @@ class OllamaConfig: timeout_seconds: float num_ctx: int seed: int + fast_model: str | None = None + fast_num_ctx: int | None = None + max_input_chars: int = 16000 class OllamaError(RuntimeError): @@ -144,13 +147,27 @@ def __init__(self, config: OllamaConfig | None = None): default_embed_model = profile.embedding_model default_num_ctx = profile.num_ctx + explicit_model = os.environ.get("CHROMEAI_OLLAMA_MODEL") + primary_model = explicit_model or default_generation_model + fast_model = os.environ.get("CHROMEAI_OLLAMA_FAST_MODEL") or ( + primary_model if explicit_model else profile.fast_generation_model + ) config = OllamaConfig( base_url=_normalize_base_url(os.environ.get("CHROMEAI_OLLAMA_BASE_URL", "http://127.0.0.1:11434")), - model=os.environ.get("CHROMEAI_OLLAMA_MODEL", default_generation_model), + model=primary_model, embed_model=os.environ.get("CHROMEAI_OLLAMA_EMBED_MODEL", default_embed_model), timeout_seconds=float(os.environ.get("CHROMEAI_OLLAMA_TIMEOUT_SECONDS", "30")), num_ctx=_positive_int(os.environ.get("CHROMEAI_OLLAMA_NUM_CTX", default_num_ctx), "CHROMEAI_OLLAMA_NUM_CTX"), seed=_nonnegative_int(os.environ.get("CHROMEAI_OLLAMA_SEED", "42"), "CHROMEAI_OLLAMA_SEED"), + fast_model=fast_model, + fast_num_ctx=_positive_int( + os.environ.get("CHROMEAI_OLLAMA_FAST_NUM_CTX", profile.fast_num_ctx), + "CHROMEAI_OLLAMA_FAST_NUM_CTX", + ), + max_input_chars=_positive_int( + os.environ.get("CHROMEAI_MAX_INPUT_CHARS", profile.max_input_chars), + "CHROMEAI_MAX_INPUT_CHARS", + ), ) self.config = config @@ -246,13 +263,30 @@ def _resolve_model(self, requested: str, preferences: list[str], models: list[st return next(model for model in models if model == requested_base or model.startswith(f"{requested_base}:")) return requested - def _generation_options(self, temperature: float) -> dict[str, Any]: + def _generation_options(self, temperature: float, num_ctx: int | None = None) -> dict[str, Any]: return { "temperature": temperature, - "num_ctx": self.config.num_ctx, + "num_ctx": num_ctx or self.config.num_ctx, "seed": self.config.seed, } + def generation_route(self, operation: str, input_text: str, local_models: list[str] | None = None) -> dict[str, Any]: + if len(input_text) > self.config.max_input_chars: + raise OllamaError( + f"selection exceeds the predictable local limit of {self.config.max_input_chars} characters" + ) + use_fast = operation in {"extract", "summarize"} + model = (self.config.fast_model or self.config.model) if use_fast else self.config.model + num_ctx = (self.config.fast_num_ctx or self.config.num_ctx) if use_fast else self.config.num_ctx + models = local_models if local_models is not None else self._model_names(local_only=True) + if not self._model_available_locally(model, models): + raise OllamaError(f"required local model is not installed: {model}") + return { + "model": model, + "num_ctx": num_ctx, + "reason": "smallest_qualified_structured_model" if use_fast else "qualified_general_model", + } + def _require_structured_object(self, raw_response: str, schema: dict[str, Any], operation: str) -> dict[str, Any]: content = _parse_jsonish(raw_response) if not isinstance(content, dict): @@ -330,13 +364,30 @@ def health(self) -> dict[str, Any]: "ignored_remote_models": [], "model_available": False, "embed_model_available": False, - "missing_models": [self.config.model, self.config.embed_model], + "missing_models": list(dict.fromkeys([ + self.config.fast_model or self.config.model, + self.config.model, + self.config.embed_model, + ])), "privacy_mode": "local-only", "hint": f"Start Ollama or point CHROMEAI_OLLAMA_BASE_URL at it.", } active_model = self.active_generation_model(local_models) active_embed_model = self.active_embedding_model(local_models) + routes = { + operation: { + "model": (self.config.fast_model or self.config.model) if operation != "agent" else self.config.model, + "num_ctx": (self.config.fast_num_ctx or self.config.num_ctx) if operation != "agent" else self.config.num_ctx, + "reason": "smallest_qualified_structured_model" if operation != "agent" else "qualified_general_model", + } + for operation in ("extract", "summarize", "agent") + } + required_models = list(dict.fromkeys([route["model"] for route in routes.values()])) + missing_generation_models = [model for model in required_models if model not in local_models] + missing_models = [*missing_generation_models, *( + [] if self.config.embed_model in local_models else [self.config.embed_model] + )] base = { "configured": True, "base_url": self.config.base_url, @@ -347,26 +398,26 @@ def health(self) -> dict[str, Any]: "timeout_seconds": self.config.timeout_seconds, "num_ctx": self.config.num_ctx, "seed": self.config.seed, + "max_input_chars": self.config.max_input_chars, + "task_routes": routes, } return { **base, "reachable": True, - "status": "ok" if active_model in local_models else "degraded", + "status": "ok" if not missing_models else "degraded", "models": all_models, "local_models": local_models, "ignored_remote_models": [name for name in all_models if name not in local_models], - "model_available": active_model in local_models, + "model_available": not missing_generation_models, "embed_model_available": active_embed_model in local_models, - "missing_models": [name for name, available in ( - (self.config.model, self.config.model in local_models), - (self.config.embed_model, self.config.embed_model in local_models), - ) if not available], + "missing_models": missing_models, "privacy_mode": "local-only", - "hint": None if active_model in local_models else f"Pull a local model such as `{self.config.model}` or set CHROMEAI_OLLAMA_MODEL to an installed local model.", + "hint": None if not missing_models else f"Install the exact required local model: {missing_models[0]}.", } def summarize(self, text: str, title: str | None = None, url: str | None = None, metadata: dict[str, Any] | None = None) -> dict[str, Any]: - model = self.active_generation_model(self._model_names(local_only=True)) + route = self.generation_route("summarize", text) + model = route["model"] schema = { "type": "object", "properties": { @@ -399,7 +450,7 @@ def summarize(self, text: str, title: str | None = None, url: str | None = None, "system": "You write precise summaries for selected text in a browser side panel.", "stream": False, "format": schema, - "options": self._generation_options(0.0), + "options": self._generation_options(0.0, route["num_ctx"]), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() @@ -421,12 +472,14 @@ def summarize(self, text: str, title: str | None = None, url: str | None = None, "model": response.get("model", model), "source": "ollama", "raw_response": raw_response, + "routing": route, } def agent(self, prompt: str, context: dict[str, Any] | None = None) -> dict[str, Any]: - model = self.active_generation_model(self._model_names(local_only=True)) context = context or {} selected_text = str(context.get("selection") or context.get("pageText") or context.get("markdown") or "").strip() + route = self.generation_route("agent", f"{prompt}\n{selected_text}".strip()) + model = route["model"] schema = { "type": "object", "properties": { @@ -454,7 +507,7 @@ def agent(self, prompt: str, context: dict[str, Any] | None = None) -> dict[str, "system": "You are a practical browser copilot that rewrites and structures selected text locally.", "stream": False, "format": schema, - "options": self._generation_options(0.0), + "options": self._generation_options(0.0, route["num_ctx"]), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() @@ -524,6 +577,7 @@ def agent(self, prompt: str, context: dict[str, Any] | None = None) -> dict[str, "model": response.get("model", model), "source": "ollama", "raw_response": raw_response, + "routing": route, } def extract( @@ -534,7 +588,8 @@ def extract( url: str | None = None, metadata: dict[str, Any] | None = None, ) -> dict[str, Any]: - model = self.active_generation_model(self._model_names(local_only=True)) + route = self.generation_route("extract", text) + model = route["model"] preset = get_extraction_preset(preset_key) prompt = ( "You are SelectPilot, a privacy-first local execution layer for selected text.\n" @@ -553,7 +608,7 @@ def extract( "system": "You generate clean structured extraction results for highlighted browser text.", "stream": False, "format": preset.schema, - "options": self._generation_options(0.0), + "options": self._generation_options(0.0, route["num_ctx"]), } response = self._request_json("/api/generate", payload) raw_response = str(response.get("response", "")).strip() @@ -586,6 +641,7 @@ def extract( "model": response.get("model", model), "source": "ollama", "raw_response": raw_response, + "routing": route, } def embed(self, text: str) -> dict[str, Any]: diff --git a/server/runtime_profiles.py b/server/runtime_profiles.py index 526b6dc..5cd0b29 100644 --- a/server/runtime_profiles.py +++ b/server/runtime_profiles.py @@ -16,8 +16,11 @@ class RuntimeProfile: label: str description: str generation_model: str + fast_generation_model: str embedding_model: str num_ctx: int + fast_num_ctx: int + max_input_chars: int target_latency: str intended_for: str is_default_auto: bool = False @@ -29,9 +32,12 @@ class RuntimeProfile: label="Fast", description="Smallest viable local profile for structured extraction and low-latency summaries.", generation_model="gemma4:e2b-it-qat", + fast_generation_model="gemma4:e2b-it-qat", embedding_model="nomic-embed-text-v2-moe:latest", num_ctx=16384, - target_latency="1-4s", + fast_num_ctx=16384, + max_input_chars=16000, + target_latency="4-20s", intended_for="Selected-text extraction, action briefs, and quick summaries.", is_default_auto=True, ), @@ -40,9 +46,12 @@ class RuntimeProfile: label="Balanced", description="Higher quality local profile for rewrite and general-purpose browser transforms.", generation_model="gemma4:e4b-it-qat", + fast_generation_model="gemma4:e2b-it-qat", embedding_model="nomic-embed-text-v2-moe:latest", num_ctx=32768, - target_latency="2-6s", + fast_num_ctx=16384, + max_input_chars=16000, + target_latency="4-30s", intended_for="Daily use when you want better quality without drifting into heavy models.", ), "advanced": RuntimeProfile( @@ -50,8 +59,11 @@ class RuntimeProfile: label="Advanced", description="Manual opt-in profile for stronger reasoning on larger machines.", generation_model="qwen2.5:7b", + fast_generation_model="gemma4:e2b-it-qat", embedding_model="nomic-embed-text-v2-moe:latest", num_ctx=32768, + fast_num_ctx=16384, + max_input_chars=16000, target_latency="4-10s", intended_for="Heavier rewrite and ask flows when latency budget is less important.", ), @@ -121,6 +133,35 @@ def list_runtime_profiles() -> list[dict[str, Any]]: return [asdict(profile) for profile in RUNTIME_PROFILES.values()] +def generation_routes(profile: RuntimeProfile) -> dict[str, dict[str, Any]]: + return { + "extract": { + "model": profile.fast_generation_model, + "num_ctx": profile.fast_num_ctx, + "reason": "smallest_qualified_structured_model", + }, + "summarize": { + "model": profile.fast_generation_model, + "num_ctx": profile.fast_num_ctx, + "reason": "smallest_qualified_structured_model", + }, + "agent": { + "model": profile.generation_model, + "num_ctx": profile.num_ctx, + "reason": "qualified_general_model", + }, + } + + +def required_generation_models(profile: RuntimeProfile) -> list[tuple[str, int]]: + required: list[tuple[str, int]] = [] + for route in generation_routes(profile).values(): + item = (str(route["model"]), int(route["num_ctx"])) + if item not in required: + required.append(item) + return required + + def get_runtime_profile(key: str | None) -> RuntimeProfile: if key in RUNTIME_PROFILES: return RUNTIME_PROFILES[key] @@ -138,6 +179,10 @@ def build_bootstrap_commands(profile_key: str, project_root: str | Path) -> dict "profile": profile.key, "command": command, "generation_model": profile.generation_model, + "fast_generation_model": profile.fast_generation_model, "embedding_model": profile.embedding_model, "num_ctx": profile.num_ctx, + "fast_num_ctx": profile.fast_num_ctx, + "max_input_chars": profile.max_input_chars, + "generation_routes": generation_routes(profile), } diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index fb31f72..49c89d7 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -18,8 +18,11 @@ test('bootstrap plan exposes exact fail-closed Fast runtime contract without sid const plan = bootstrapPlan('fast'); assert.equal(plan.selected_profile, 'fast'); assert.equal(plan.generation_model, 'gemma4:e2b-it-qat'); + assert.equal(plan.fast_generation_model, 'gemma4:e2b-it-qat'); assert.equal(plan.embedding_model, 'nomic-embed-text-v2-moe:latest'); assert.equal(plan.num_ctx, 16_384); + assert.equal(plan.fast_num_ctx, 16_384); + assert.equal(plan.max_input_chars, 16_000); assert.equal(plan.reason, 'Explicit profile selected by operator.'); }); @@ -27,8 +30,15 @@ test('bootstrap plan exposes exact fail-closed Balanced runtime contract without const plan = bootstrapPlan('balanced'); assert.equal(plan.selected_profile, 'balanced'); assert.equal(plan.generation_model, 'gemma4:e4b-it-qat'); + assert.equal(plan.fast_generation_model, 'gemma4:e2b-it-qat'); assert.equal(plan.embedding_model, 'nomic-embed-text-v2-moe:latest'); assert.equal(plan.num_ctx, 32_768); + assert.equal(plan.fast_num_ctx, 16_384); + assert.deepEqual(plan.generation_routes, { + extract: { model: 'gemma4:e2b-it-qat', num_ctx: 16_384, reason: 'smallest_qualified_structured_model' }, + summarize: { model: 'gemma4:e2b-it-qat', num_ctx: 16_384, reason: 'smallest_qualified_structured_model' }, + agent: { model: 'gemma4:e4b-it-qat', num_ctx: 32_768, reason: 'qualified_general_model' }, + }); assert.equal(plan.reason, 'Explicit profile selected by operator.'); }); diff --git a/tests/panel/runtime-policy-determinism.test.mjs b/tests/panel/runtime-policy-determinism.test.mjs new file mode 100644 index 0000000..1609da7 --- /dev/null +++ b/tests/panel/runtime-policy-determinism.test.mjs @@ -0,0 +1,19 @@ +// module_name: runtime_policy_determinism_test +// spec_ref: "prompt_determinism" + +import assert from 'node:assert/strict'; +import test from 'node:test'; + +import { replaceCompileAuditEvent } from '../../scripts/runtime-policy-lib.mjs'; + +test('policy compilation keeps one deterministic compile audit event', () => { + const historical = { event_type: 'rollback', model_id: 'historical:model' }; + const staleCompile = { event_type: 'compile_policy', policy_version: 'old' }; + const currentCompile = { event_type: 'compile_policy', policy_version: 'current' }; + + const once = replaceCompileAuditEvent([historical, staleCompile], currentCompile); + const twice = replaceCompileAuditEvent(once, currentCompile); + + assert.deepEqual(twice, once); + assert.deepEqual(twice, [historical, currentCompile]); +}); diff --git a/tests/server/test_ollama_client.py b/tests/server/test_ollama_client.py index 9fe2fb2..c622c2a 100644 --- a/tests/server/test_ollama_client.py +++ b/tests/server/test_ollama_client.py @@ -15,20 +15,27 @@ class StubOllamaClient(OllamaClient): - def __init__(self, response: dict): + def __init__(self, response: dict, *, balanced: bool = False): super().__init__(OllamaConfig( base_url="http://127.0.0.1:11434", - model="gemma4:e2b-it-qat", + model="gemma4:e4b-it-qat" if balanced else "gemma4:e2b-it-qat", embed_model="nomic-embed-text-v2-moe:latest", timeout_seconds=1, - num_ctx=16384, + num_ctx=32768 if balanced else 16384, seed=42, + fast_model="gemma4:e2b-it-qat", + fast_num_ctx=16384, + max_input_chars=16000, )) self.response = response self.last_payload = None def _model_names(self, local_only: bool = False) -> list[str]: - return [self.config.model, self.config.embed_model] + return list(dict.fromkeys([ + self.config.fast_model, + self.config.model, + self.config.embed_model, + ])) def _request_json(self, path: str, payload=None): self.last_payload = payload @@ -53,6 +60,32 @@ def test_summarize_sends_explicit_context_window(self) -> None: self.assertEqual(client.last_payload["options"]["seed"], 42) self.assertEqual(client.last_payload["options"]["temperature"], 0.0) + def test_balanced_routes_structured_and_agent_tasks_to_exact_models(self) -> None: + client = StubOllamaClient({ + "model": "gemma4:e2b-it-qat", + "response": '{"summary":"S","bullets":[],"action_items":[],"title":"T","tags":[]}', + }, balanced=True) + summary = client.summarize("content") + self.assertEqual(client.last_payload["model"], "gemma4:e2b-it-qat") + self.assertEqual(client.last_payload["options"]["num_ctx"], 16384) + self.assertEqual(summary["routing"]["reason"], "smallest_qualified_structured_model") + + client.response = {"model": "gemma4:e4b-it-qat", "response": '{"reasoning":[],"markdown":"M","json":{}}'} + agent = client.agent("prompt") + self.assertEqual(client.last_payload["model"], "gemma4:e4b-it-qat") + self.assertEqual(client.last_payload["options"]["num_ctx"], 32768) + self.assertEqual(agent["routing"]["reason"], "qualified_general_model") + + def test_missing_routed_model_fails_without_substitution(self) -> None: + client = StubOllamaClient({}, balanced=True) + with self.assertRaisesRegex(OllamaError, "required local model is not installed: gemma4:e2b-it-qat"): + client.generation_route("extract", "content", ["gemma4:e4b-it-qat"]) + + def test_long_selection_fails_before_inference(self) -> None: + client = StubOllamaClient({}, balanced=True) + with self.assertRaisesRegex(OllamaError, "predictable local limit"): + client.generation_route("extract", "x" * 16001, ["gemma4:e2b-it-qat"]) + def test_summarize_rejects_non_schema_output(self) -> None: client = StubOllamaClient({"response": "not json"}) with self.assertRaisesRegex(OllamaError, "non-object structured output"): diff --git a/tests/server/test_runtime_profiles.py b/tests/server/test_runtime_profiles.py index 733de38..78835bf 100644 --- a/tests/server/test_runtime_profiles.py +++ b/tests/server/test_runtime_profiles.py @@ -12,7 +12,13 @@ if str(SERVER_DIR) not in sys.path: sys.path.insert(0, str(SERVER_DIR)) -from runtime_profiles import build_bootstrap_commands, get_runtime_profile, recommend_runtime_profile # noqa: E402 +from runtime_profiles import ( # noqa: E402 + build_bootstrap_commands, + generation_routes, + get_runtime_profile, + recommend_runtime_profile, + required_generation_models, +) from ollama_client import OllamaClient, OllamaError # noqa: E402 @@ -50,6 +56,7 @@ def test_explicit_profile_and_context_override_are_authoritative(self) -> None: }, clear=True): client = OllamaClient() self.assertEqual(client.config.model, "custom-local:model") + self.assertEqual(client.config.fast_model, "custom-local:model") self.assertEqual(client.config.num_ctx, 24576) def test_invalid_context_override_fails_explicitly(self) -> None: @@ -61,6 +68,19 @@ def test_bootstrap_command_contains_profile(self) -> None: command = build_bootstrap_commands("balanced", ROOT) self.assertIn("--profile balanced", command["command"]) self.assertEqual(command["num_ctx"], 32768) + self.assertEqual(command["fast_num_ctx"], 16384) + self.assertEqual(command["max_input_chars"], 16000) + + def test_balanced_profile_routes_smallest_qualified_model_per_task(self) -> None: + profile = get_runtime_profile("balanced") + routes = generation_routes(profile) + self.assertEqual(routes["extract"]["model"], "gemma4:e2b-it-qat") + self.assertEqual(routes["summarize"]["num_ctx"], 16384) + self.assertEqual(routes["agent"]["model"], "gemma4:e4b-it-qat") + self.assertEqual(required_generation_models(profile), [ + ("gemma4:e2b-it-qat", 16384), + ("gemma4:e4b-it-qat", 32768), + ]) if __name__ == "__main__": diff --git a/tests/server/test_validation_pipeline.py b/tests/server/test_validation_pipeline.py index f369205..c977185 100644 --- a/tests/server/test_validation_pipeline.py +++ b/tests/server/test_validation_pipeline.py @@ -5,6 +5,7 @@ import sys import unittest from pathlib import Path +from types import SimpleNamespace from unittest.mock import patch ROOT = Path(__file__).resolve().parents[2] @@ -99,6 +100,14 @@ def test_extract_payload_rejects_non_object_metadata(self) -> None: self.assertEqual(ctx.exception.code, "invalid_request_field") self.assertEqual(ctx.exception.details.get("field"), "metadata") + def test_long_selection_is_rejected_before_inference(self) -> None: + with patch("nano_server.OLLAMA", SimpleNamespace(config=SimpleNamespace(max_input_chars=10))): + with self.assertRaises(ValidationError) as ctx: + validate_extract_payload({"text": "x" * 11}) + self.assertEqual(ctx.exception.code, "selection_too_large") + self.assertEqual(ctx.exception.status, 413) + self.assertEqual(ctx.exception.details, {"maximum_characters": 10, "received_characters": 11}) + def test_agent_payload_requires_context_object_when_present(self) -> None: with self.assertRaises(ValidationError) as ctx: validate_agent_payload({"prompt": "go", "context": "bad"}) @@ -122,6 +131,7 @@ def test_summarize_response_requires_schema(self) -> None: "model": "qwen", "source": "ollama", "raw_response": "{}", + "routing": {"model": "qwen", "num_ctx": 16384, "reason": "test_route"}, } result = validate_summarize_response(valid) self.assertEqual(result["source"], "ollama") @@ -139,6 +149,7 @@ def test_agent_response_requires_reasoning_and_json_object(self) -> None: "model": "qwen", "source": "ollama", "raw_response": "{}", + "routing": {"model": "qwen", "num_ctx": 16384, "reason": "test_route"}, } result = validate_agent_response(valid) self.assertEqual(result["model"], "qwen") @@ -156,6 +167,7 @@ def test_extract_response_requires_json_object(self) -> None: "model": "qwen", "source": "ollama", "raw_response": "{}", + "routing": {"model": "qwen", "num_ctx": 16384, "reason": "test_route"}, } result = validate_extract_response(valid) self.assertEqual(result["preset"], "action_brief") @@ -280,6 +292,7 @@ def execute_fn(): "model": "qwen", "source": "ollama", "raw_response": "{}", + "routing": {"model": "qwen", "num_ctx": 16384, "reason": "test_route"}, } return { "summary": "ok", @@ -291,6 +304,7 @@ def execute_fn(): "model": "qwen", "source": "ollama", "raw_response": "{}", + "routing": {"model": "qwen", "num_ctx": 16384, "reason": "test_route"}, } with patch("nano_server.emit_runtime_meta") as emit_mock: From f46662d6232353a882e27fc51e2fa094b3cbe8f2 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 19:35:13 +0200 Subject: [PATCH 43/66] fix(ci): verify missing models fail closed Replace the stale model-downgrade stress premise with explicit unavailable-model evidence, no substitution, and zero external calls.\n\nVerified: node scripts/stress-runner.mjs; pnpm lint; pnpm verify:admission. --- reports/stress/master_summary.json | 14 +++--- reports/stress/selectpilot_monolith.json | 15 +++--- reports/stress/stress_test_cases.json | 9 ++-- .../stress/tests/T001_clean_extraction.json | 24 ++++----- reports/stress/tests/T002_ambiguity_gate.json | 24 ++++----- .../tests/T003_contract_enforcement.json | 24 ++++----- .../stress/tests/T004_conflicting_input.json | 24 ++++----- reports/stress/tests/T005_large_context.json | 24 ++++----- .../stress/tests/T006_runtime_mismatch.json | 49 +++++++++---------- .../tests/T007_retrieval_pollution.json | 24 ++++----- .../stress/tests/T008_privacy_boundary.json | 24 ++++----- .../stress/tests/T009_parallel_execution.json | 24 ++++----- reports/stress/tests/T010_recovery_path.json | 24 ++++----- scripts/stress-runner.mjs | 12 ++++- 14 files changed, 161 insertions(+), 154 deletions(-) diff --git a/reports/stress/master_summary.json b/reports/stress/master_summary.json index 18e1b9a..020570d 100644 --- a/reports/stress/master_summary.json +++ b/reports/stress/master_summary.json @@ -1,14 +1,14 @@ { "suite_id": "artificial_diagnostic_load_suite_v1", "entity_id": "balanced_entity_v1", - "overall_score": 0.9078, + "overall_score": 0.9064, "band": "robust", "critical_failures": 0, "high_failures": 0, "dominant_weak_points": [ - "runtime_adaptivity", "recovery_capacity", - "semantic_precision" + "semantic_precision", + "resource_efficiency" ], "dominant_strengths": [ "privacy_boundary_integrity", @@ -53,10 +53,10 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.92, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, - "runtime_adaptivity": 0.846, - "recovery_capacity": 0.83, + "runtime_adaptivity": 0.84, + "recovery_capacity": 0.82, "output_consistency": 0.95 }, "targets": { @@ -95,7 +95,7 @@ "bounded_fallback": true, "file_backed_reproducibility": true }, - "generated_at_unix_ms": 1703418057844, + "generated_at_unix_ms": 1703325128577, "required_phases": [ "baseline_scan", "structural_ct", diff --git a/reports/stress/selectpilot_monolith.json b/reports/stress/selectpilot_monolith.json index 0954690..f115cfc 100644 --- a/reports/stress/selectpilot_monolith.json +++ b/reports/stress/selectpilot_monolith.json @@ -326,13 +326,14 @@ "content": "Extract JSON under constrained model availability." }, "expected": { - "fallback_triggered": true, - "runtime_profile": "downgraded", - "no_silent_failure": true + "execution_blocked": true, + "model_substitution": false, + "failure_mode": "explicit" }, "fail_conditions": [ "crash", - "no_fallback", + "execution_continued", + "implicit_fallback", "silent_model_switch" ] }, @@ -480,9 +481,7 @@ "baseline_selector", "baseline_smallest_sufficient_selection_from_registry" ], - "fallback_selection_paths": [ - "runtime_policy_fallback" - ], + "fallback_selection_paths": [], "resource_efficiency_small_model_pattern": "0.5b" } }, @@ -646,4 +645,4 @@ "strict_form": "Execution may deform within constraints, but must always return to an identical pre-execution state with no residual effects." } } -} \ No newline at end of file +} diff --git a/reports/stress/stress_test_cases.json b/reports/stress/stress_test_cases.json index 64ce158..a962c5d 100644 --- a/reports/stress/stress_test_cases.json +++ b/reports/stress/stress_test_cases.json @@ -118,13 +118,14 @@ "content": "Extract JSON under constrained model availability." }, "expected": { - "fallback_triggered": true, - "runtime_profile": "downgraded", - "no_silent_failure": true + "execution_blocked": true, + "model_substitution": false, + "failure_mode": "explicit" }, "fail_conditions": [ "crash", - "no_fallback", + "execution_continued", + "implicit_fallback", "silent_model_switch" ] }, diff --git a/reports/stress/tests/T001_clean_extraction.json b/reports/stress/tests/T001_clean_extraction.json index cbf6fea..abcb153 100644 --- a/reports/stress/tests/T001_clean_extraction.json +++ b/reports/stress/tests/T001_clean_extraction.json @@ -73,24 +73,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_431776f13178edc3", - "pipeline_trace_b": "sp_431776f13178edc3", + "pipeline_trace_a": "sp_6186e96f8cdb41e2", + "pipeline_trace_b": "sp_6186e96f8cdb41e2", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -167,7 +167,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -237,7 +237,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -249,7 +249,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T002_ambiguity_gate.json b/reports/stress/tests/T002_ambiguity_gate.json index 58c9b9f..5530897 100644 --- a/reports/stress/tests/T002_ambiguity_gate.json +++ b/reports/stress/tests/T002_ambiguity_gate.json @@ -70,24 +70,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_87dab4b77df5191f", - "pipeline_trace_b": "sp_87dab4b77df5191f", + "pipeline_trace_a": "sp_feafaaa93d0efb7e", + "pipeline_trace_b": "sp_feafaaa93d0efb7e", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -164,7 +164,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -234,7 +234,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "pass", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -246,7 +246,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 1, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T003_contract_enforcement.json b/reports/stress/tests/T003_contract_enforcement.json index dede15a..10c05b8 100644 --- a/reports/stress/tests/T003_contract_enforcement.json +++ b/reports/stress/tests/T003_contract_enforcement.json @@ -67,24 +67,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_1da3bc412de9046d", - "pipeline_trace_b": "sp_1da3bc412de9046d", + "pipeline_trace_a": "sp_ed4ea9dfb17c0b9c", + "pipeline_trace_b": "sp_ed4ea9dfb17c0b9c", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -161,7 +161,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -231,7 +231,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -243,7 +243,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T004_conflicting_input.json b/reports/stress/tests/T004_conflicting_input.json index 404fa56..772ce4f 100644 --- a/reports/stress/tests/T004_conflicting_input.json +++ b/reports/stress/tests/T004_conflicting_input.json @@ -70,24 +70,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_7f8aee399feeccbf", - "pipeline_trace_b": "sp_7f8aee399feeccbf", + "pipeline_trace_a": "sp_090dfd17f5fd97a2", + "pipeline_trace_b": "sp_090dfd17f5fd97a2", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -164,7 +164,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -234,7 +234,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "pass", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -246,7 +246,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 1, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T005_large_context.json b/reports/stress/tests/T005_large_context.json index 5efbc80..56d0401 100644 --- a/reports/stress/tests/T005_large_context.json +++ b/reports/stress/tests/T005_large_context.json @@ -63,24 +63,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e2b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_620acf79e1b90d84", - "pipeline_trace_b": "sp_620acf79e1b90d84", + "pipeline_trace_a": "sp_e46a7ad798bf3b21", + "pipeline_trace_b": "sp_e46a7ad798bf3b21", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -157,7 +157,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e2b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -227,7 +227,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -239,7 +239,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T006_runtime_mismatch.json b/reports/stress/tests/T006_runtime_mismatch.json index 0fbc506..64eb574 100644 --- a/reports/stress/tests/T006_runtime_mismatch.json +++ b/reports/stress/tests/T006_runtime_mismatch.json @@ -4,13 +4,14 @@ "phase": "trauma_simulation", "input_profile": "normal_request", "expected": { - "fallback_triggered": true, - "runtime_profile": "downgraded", - "no_silent_failure": true + "execution_blocked": true, + "model_substitution": false, + "failure_mode": "explicit" }, "declared_fail_conditions": [ "crash", - "no_fallback", + "execution_continued", + "implicit_fallback", "silent_model_switch" ], "expected_runtime_mismatch_allowed": true, @@ -63,24 +64,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", - "selection_path": "runtime_policy_fallback", - "selection_reason": "runtime_policy_fallback_no_models_available_simulated_degraded_mode", - "policy_version": "runtime-policy-bbc138feed76", + "selected_model_id": null, + "selection_path": "explicit_model_unavailable", + "selection_reason": "required_model_missing_fail_closed", + "policy_version": "runtime-policy-8226a11e6ae3", "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": null, "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", - "selection_path": "runtime_policy_fallback", - "selection_reason": "runtime_policy_fallback_no_models_available_simulated_degraded_mode", - "policy_version": "runtime-policy-bbc138feed76", + "selected_model_id": null, + "selection_path": "explicit_model_unavailable", + "selection_reason": "required_model_missing_fail_closed", + "policy_version": "runtime-policy-8226a11e6ae3", "promotion_applied": false }, - "pipeline_trace_a": "sp_eed3e21ecb6df3ca", - "pipeline_trace_b": "sp_eed3e21ecb6df3ca", + "pipeline_trace_a": "sp_bd48c6988b95bae3", + "pipeline_trace_b": "sp_bd48c6988b95bae3", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -156,8 +157,8 @@ "prompt_version": "deterministic.prompts.v3" } }, - "runtime_selection": "runtime_policy_fallback", - "active_model": "qwen2.5:0.5b", + "runtime_selection": "explicit_model_unavailable", + "active_model": null, "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -165,9 +166,7 @@ "memory_writes": [ "runtime/live_feedback.jsonl" ], - "fallback_events": [ - "fallback_or_baseline_path_activated" - ], + "fallback_events": [], "violation_events": [], "latency_profile": { "budget_ms": 2000 @@ -229,7 +228,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -241,13 +240,13 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, - "runtime_adaptivity": 0.9, - "recovery_capacity": 0.92, + "runtime_adaptivity": 0.84, + "recovery_capacity": 0.82, "output_consistency": 0.95, "retrieval_precision": 0.78, - "overall_score": 0.922 + "overall_score": 0.908 }, "pass": true } diff --git a/reports/stress/tests/T007_retrieval_pollution.json b/reports/stress/tests/T007_retrieval_pollution.json index 271b60b..8645a4e 100644 --- a/reports/stress/tests/T007_retrieval_pollution.json +++ b/reports/stress/tests/T007_retrieval_pollution.json @@ -69,24 +69,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_824d98e4b6fa060b", - "pipeline_trace_b": "sp_824d98e4b6fa060b", + "pipeline_trace_a": "sp_25cb6e9a24d96a7a", + "pipeline_trace_b": "sp_25cb6e9a24d96a7a", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -163,7 +163,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -233,7 +233,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -245,7 +245,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T008_privacy_boundary.json b/reports/stress/tests/T008_privacy_boundary.json index 5e7d0cd..6629c4e 100644 --- a/reports/stress/tests/T008_privacy_boundary.json +++ b/reports/stress/tests/T008_privacy_boundary.json @@ -69,24 +69,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_ef61357d7ea1bae2", - "pipeline_trace_b": "sp_ef61357d7ea1bae2", + "pipeline_trace_a": "sp_13cf094342f08a8b", + "pipeline_trace_b": "sp_13cf094342f08a8b", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -163,7 +163,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -233,7 +233,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -245,7 +245,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T009_parallel_execution.json b/reports/stress/tests/T009_parallel_execution.json index e0c4c83..e5b6176 100644 --- a/reports/stress/tests/T009_parallel_execution.json +++ b/reports/stress/tests/T009_parallel_execution.json @@ -71,24 +71,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_4b8c9d891186235f", - "pipeline_trace_b": "sp_4b8c9d891186235f", + "pipeline_trace_a": "sp_3bc199c7c185f2be", + "pipeline_trace_b": "sp_3bc199c7c185f2be", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -165,7 +165,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -235,7 +235,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -247,7 +247,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/reports/stress/tests/T010_recovery_path.json b/reports/stress/tests/T010_recovery_path.json index 6515fc6..5049e8a 100644 --- a/reports/stress/tests/T010_recovery_path.json +++ b/reports/stress/tests/T010_recovery_path.json @@ -69,24 +69,24 @@ } }, "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "external_calls": 0, "observed_behavior": { "runtime_selection": { - "selected_model_id": "qwen2.5:0.5b", + "selected_model_id": "gemma4:e4b-it-qat", "selection_path": "runtime_policy_preferred", "selection_reason": "baseline_registry_no_runtime_promotion", - "policy_version": "runtime-policy-bbc138feed76", - "promotion_applied": true + "policy_version": "runtime-policy-8226a11e6ae3", + "promotion_applied": false }, - "pipeline_trace_a": "sp_5c3cc2e85da13e0a", - "pipeline_trace_b": "sp_5c3cc2e85da13e0a", + "pipeline_trace_a": "sp_b79b89f6d83d78b7", + "pipeline_trace_b": "sp_b79b89f6d83d78b7", "deterministic_repeat_stable": true, "state_reset_checks": { "ephemeral_state_cleared": true, @@ -163,7 +163,7 @@ } }, "runtime_selection": "runtime_policy_preferred", - "active_model": "qwen2.5:0.5b", + "active_model": "gemma4:e4b-it-qat", "memory_reads": [ "runtime/model_policy.json", "runtime/model_registry.runtime.json" @@ -233,7 +233,7 @@ "memory_integrity": "warning", "contract_obedience": "pass", "ambiguity_handling": "warning", - "resource_efficiency": "warning", + "resource_efficiency": "fail", "privacy_boundary_integrity": "pass", "runtime_adaptivity": "warning", "recovery_capacity": "warning", @@ -245,7 +245,7 @@ "memory_integrity": 0.92, "contract_obedience": 0.98, "ambiguity_handling": 0.9, - "resource_efficiency": 0.9, + "resource_efficiency": 0.75, "privacy_boundary_integrity": 1, "runtime_adaptivity": 0.84, "recovery_capacity": 0.82, diff --git a/scripts/stress-runner.mjs b/scripts/stress-runner.mjs index d31b6c6..f081d44 100644 --- a/scripts/stress-runner.mjs +++ b/scripts/stress-runner.mjs @@ -436,7 +436,7 @@ async function runSingleTest(test, monolith, bindings, runtimePolicy, runtimeReg ? installedModelIds.filter((modelId) => modelId !== matchingPolicyTuple?.preferred_model_id) : installedModelIds; - const selected = bindings.selectRuntimeModel( + let selected = bindings.selectRuntimeModel( { taskFamily: taskAnalysis.task_family, outputMode: taskAnalysis.output_mode, @@ -446,7 +446,15 @@ async function runSingleTest(test, monolith, bindings, runtimePolicy, runtimeReg runtimePolicy, runtimeRegistry, ); - if (!selected) { + if (!selected && expectedRuntimeMismatchAllowed) { + selected = { + selected_model_id: null, + selection_path: 'explicit_model_unavailable', + selection_reason: 'required_model_missing_fail_closed', + policy_version: runtimePolicy.policy_version, + promotion_applied: false, + }; + } else if (!selected) { throw new Error('runtime_policy_no_match'); } From b429b2f1e073f7f6d8180def6b808f3bb2979419 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 22:11:42 +0200 Subject: [PATCH 44/66] fix(macos): install persistent local runtime Install the launchd runtime outside privacy-protected source folders, require Python 3.10 or newer for both local and packaged helpers, and wait for bridge readiness. Verified: pnpm lint; pnpm test; pnpm build; pnpm verify:admission; pnpm test:privacy; pnpm test:billing; pnpm lint:manifest; pnpm validate:store; pnpm test:e2e --- .../com.selectpilot.bridge.plist.template | 2 +- installer/macos/scripts/postinstall | 15 ++++++- launchd/com.chromeai.nano.plist | 8 ++-- scripts/bootstrap-macos-local.sh | 9 ++++- scripts/install-macos-local.sh | 37 ++++++++++++++++-- tests/panel/bootstrap-contract.test.mjs | 39 +++++++++++++++++++ 6 files changed, 100 insertions(+), 10 deletions(-) diff --git a/installer/macos/com.selectpilot.bridge.plist.template b/installer/macos/com.selectpilot.bridge.plist.template index 4d8f375..582c9d8 100644 --- a/installer/macos/com.selectpilot.bridge.plist.template +++ b/installer/macos/com.selectpilot.bridge.plist.template @@ -3,7 +3,7 @@ Labelcom.selectpilot.bridge ProgramArguments - /usr/bin/python3__PROJECT_ROOT__/server/nano_server.py + __PYTHON_BIN____PROJECT_ROOT__/server/nano_server.py --port8083 --run-dir__RUN_DIR__ --log-dir__LOG_DIR__ diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall index 1408c7b..2fa2b94 100644 --- a/installer/macos/scripts/postinstall +++ b/installer/macos/scripts/postinstall @@ -15,7 +15,19 @@ LOG_DIR="$USER_HOME/Library/Logs/SelectPilot" AGENT_DIR="$USER_HOME/Library/LaunchAgents" AGENT_PATH="$AGENT_DIR/com.selectpilot.bridge.plist" HASH="$(shasum -a 256 "$SOURCE_ROOT/server/nano_server.py" | awk '{print $1}')" -PROFILE_VALUES="$(sudo -u "$CONSOLE_USER" /usr/bin/python3 - "$SOURCE_ROOT/server" <<'PY' +PYTHON_BIN="" +for candidate in /opt/homebrew/bin/python3 /usr/local/bin/python3 /Library/Frameworks/Python.framework/Versions/Current/bin/python3 /usr/bin/python3; do + if [ -x "$candidate" ] && "$candidate" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, 10) else 1)'; then + PYTHON_BIN="$candidate" + break + fi +done +if [ -z "$PYTHON_BIN" ]; then + echo "SelectPilot requires Python 3.10 or newer." >&2 + exit 1 +fi + +PROFILE_VALUES="$(sudo -u "$CONSOLE_USER" "$PYTHON_BIN" - "$SOURCE_ROOT/server" <<'PY' import sys sys.path.insert(0, sys.argv[1]) from runtime_profiles import get_runtime_profile, recommend_runtime_profile @@ -29,6 +41,7 @@ EOF install -d -m 700 -o "$CONSOLE_USER" "$RUN_DIR" "$LOG_DIR" "$AGENT_DIR" sed \ + -e "s|__PYTHON_BIN__|$PYTHON_BIN|g" \ -e "s|__PROJECT_ROOT__|$SOURCE_ROOT|g" \ -e "s|__RUN_DIR__|$RUN_DIR|g" \ -e "s|__LOG_DIR__|$LOG_DIR|g" \ diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index 4b4de4e..6415fbe 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -5,18 +5,18 @@ Labelcom.chromeai.nano ProgramArguments - /usr/bin/python3 - __PROJECT_ROOT__/server/nano_server.py + __PYTHON_BIN__ + __INSTALLED_BINARY__ --port8083 --run-dir__RUN_DIR__ --log-dir__LOG_DIR__ - --binary-path__PROJECT_ROOT__/server/nano_server.py + --binary-path__INSTALLED_BINARY__ RunAtLoad KeepAliveSuccessfulExit StandardOutPath__LOG_DIR__/nano.log StandardErrorPath__LOG_DIR__/nano.err - WorkingDirectory__PROJECT_ROOT__ + WorkingDirectory__INSTALL_DIR__ EnvironmentVariables CHROMEAI_PORT_INFO__RUN_DIR__/port.info diff --git a/scripts/bootstrap-macos-local.sh b/scripts/bootstrap-macos-local.sh index 70657a3..53bb91e 100755 --- a/scripts/bootstrap-macos-local.sh +++ b/scripts/bootstrap-macos-local.sh @@ -205,7 +205,14 @@ CHROMEAI_OLLAMA_SEED=42 \ "$ROOT/scripts/install-macos-local.sh" STATUS_LAUNCHAGENT="ok" -HEALTH_JSON="$(curl -sSf "$BRIDGE_HEALTH_URL" || true)" +HEALTH_JSON="" +for _attempt in $(seq 1 15); do + HEALTH_JSON="$(curl -sSf "$BRIDGE_HEALTH_URL" 2>/dev/null || true)" + if [[ -n "$HEALTH_JSON" ]]; then + break + fi + sleep 1 +done if python3 - "$HEALTH_JSON" <<'PY' import json import sys diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index 36fe2c3..5f2da3e 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -4,7 +4,16 @@ set -euo pipefail ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" TEMPLATE="$ROOT/launchd/com.chromeai.nano.plist" DEST="${HOME}/Library/LaunchAgents/com.chromeai.nano.plist" -HASH="$(shasum -a 256 "$ROOT/server/nano_server.py" | awk '{print $1}')" +APP_DIR="${CHROMEAI_APP_DIR:-${HOME}/Library/Application Support/SelectPilot}" +INSTALL_DIR="$APP_DIR/server" +SOURCE_BINARY="$ROOT/server/nano_server.py" +INSTALLED_BINARY="$INSTALL_DIR/nano_server.py" +RUNTIME_MODULES=( + nano_server.py + ollama_client.py + extraction_presets.py + runtime_profiles.py +) OLLAMA_BASE_URL="${CHROMEAI_OLLAMA_BASE_URL:-http://127.0.0.1:11434}" OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-gemma4:e2b-it-qat}" OLLAMA_FAST_MODEL="${CHROMEAI_OLLAMA_FAST_MODEL:-$OLLAMA_MODEL}" @@ -15,12 +24,32 @@ MAX_INPUT_CHARS="${CHROMEAI_MAX_INPUT_CHARS:-16000}" OLLAMA_SEED="${CHROMEAI_OLLAMA_SEED:-42}" RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run}" LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}" +PYTHON_BIN="${CHROMEAI_PYTHON_BIN:-$(command -v python3)}" -mkdir -p "${HOME}/Library/LaunchAgents" +if ! "$PYTHON_BIN" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, 10) else 1)'; then + echo "SelectPilot requires Python 3.10 or newer; found $($PYTHON_BIN --version 2>&1)." >&2 + exit 1 +fi + +mkdir -p "${HOME}/Library/LaunchAgents" "$INSTALL_DIR" "$APP_DIR/presets" "$APP_DIR/runtime" mkdir -p "$RUN_DIR" "$LOG_DIR" +# LaunchAgents cannot reliably read repositories located in macOS privacy- +# protected folders such as Documents. Install an immutable runtime copy in +# Application Support and bind the integrity contract to that exact copy. +for module in "${RUNTIME_MODULES[@]}"; do + install -m 0555 "$ROOT/server/$module" "$INSTALL_DIR/$module" +done +install -m 0444 "$ROOT/presets/extraction-presets.json" "$APP_DIR/presets/extraction-presets.json" +for policy in model_policy.json model_registry.runtime.json promotion_audit.json; do + install -m 0444 "$ROOT/runtime/$policy" "$APP_DIR/runtime/$policy" +done +HASH="$(shasum -a 256 "$INSTALLED_BINARY" | awk '{print $1}')" + sed \ - -e "s|__PROJECT_ROOT__|$ROOT|g" \ + -e "s|__PYTHON_BIN__|$PYTHON_BIN|g" \ + -e "s|__INSTALL_DIR__|$INSTALL_DIR|g" \ + -e "s|__INSTALLED_BINARY__|$INSTALLED_BINARY|g" \ -e "s|__RUN_DIR__|$RUN_DIR|g" \ -e "s|__LOG_DIR__|$LOG_DIR|g" \ -e "s|__BINARY_HASH__|$HASH|g" \ @@ -40,6 +69,8 @@ launchctl load "$DEST" cat < { + const bootstrap = readFileSync('./scripts/bootstrap-macos-local.sh', 'utf8'); + + assert.match(bootstrap, /for _attempt in \$\(seq 1 15\)/); + assert.match(bootstrap, /curl -sSf "\$BRIDGE_HEALTH_URL"/); + assert.match(bootstrap, /sleep 1/); +}); + +test('macOS service installs outside the source checkout before launchd execution', () => { + const installer = readFileSync('./scripts/install-macos-local.sh', 'utf8'); + const launchAgent = readFileSync('./launchd/com.chromeai.nano.plist', 'utf8'); + + assert.match(installer, /Application Support\/SelectPilot/); + assert.match(installer, /RUNTIME_MODULES=\(/); + assert.match(installer, /ollama_client\.py/); + assert.match(installer, /extraction_presets\.py/); + assert.match(installer, /runtime_profiles\.py/); + assert.match(installer, /presets\/extraction-presets\.json/); + assert.match(installer, /model_policy\.json/); + assert.match(installer, /model_registry\.runtime\.json/); + assert.match(installer, /sys\.version_info >= \(3, 10\)/); + assert.match(launchAgent, /__PYTHON_BIN__/); + assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); + assert.match(installer, /install -m 0555 "\$ROOT\/server\/\$module" "\$INSTALL_DIR\/\$module"/); + assert.match(launchAgent, /__INSTALLED_BINARY__/); + assert.match(launchAgent, /__INSTALL_DIR__/); + assert.doesNotMatch(launchAgent, /__PROJECT_ROOT__/); +}); + +test('production macOS package also binds launchd to Python 3.10 or newer', () => { + const postinstall = readFileSync('./installer/macos/scripts/postinstall', 'utf8'); + const launchAgent = readFileSync('./installer/macos/com.selectpilot.bridge.plist.template', 'utf8'); + + assert.match(postinstall, /sys\.version_info >= \(3, 10\)/); + assert.match(postinstall, /s\|__PYTHON_BIN__\|\$PYTHON_BIN\|g/); + assert.match(launchAgent, /__PYTHON_BIN__/); + assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); +}); From 85e52161925e2ea8789a275f79db7c8ce486be0c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 22:14:44 +0200 Subject: [PATCH 45/66] fix(ci): remove obsolete installer variable Verified: shellcheck scripts/*.sh; pnpm lint; pnpm test --- scripts/install-macos-local.sh | 1 - 1 file changed, 1 deletion(-) diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index 5f2da3e..d29e79f 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -6,7 +6,6 @@ TEMPLATE="$ROOT/launchd/com.chromeai.nano.plist" DEST="${HOME}/Library/LaunchAgents/com.chromeai.nano.plist" APP_DIR="${CHROMEAI_APP_DIR:-${HOME}/Library/Application Support/SelectPilot}" INSTALL_DIR="$APP_DIR/server" -SOURCE_BINARY="$ROOT/server/nano_server.py" INSTALLED_BINARY="$INSTALL_DIR/nano_server.py" RUNTIME_MODULES=( nano_server.py From c1d738e8a7c95082bef9af60a6d67ebdeedeae7d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 22:25:00 +0200 Subject: [PATCH 46/66] fix(macos): support system Python 3.9 Align both macOS installation paths with the declared Python 3.9 compatibility contract and keep the minimum version explicit in regression coverage. Verified: shellcheck scripts/install-macos-local.sh installer/macos/scripts/postinstall; node --test tests/panel/bootstrap-contract.test.mjs; full Hetzner release verification. --- CONTEXT.md | 10 ++++++++++ installer/macos/scripts/postinstall | 4 ++-- scripts/install-macos-local.sh | 4 ++-- tests/panel/bootstrap-contract.test.mjs | 6 +++--- 4 files changed, 17 insertions(+), 7 deletions(-) create mode 100644 CONTEXT.md diff --git a/CONTEXT.md b/CONTEXT.md new file mode 100644 index 0000000..37c1617 --- /dev/null +++ b/CONTEXT.md @@ -0,0 +1,10 @@ +# SelectPilot Product Context + +This context defines stable product and platform terms used by SelectPilot release contracts. + +## Language + +**Supported macOS Python runtime**: +Python 3.9 or newer, matching the oldest declared system runtime on supported Apple hardware. +_Avoid_: Python 3.10 minimum +_Authority_: Linear SOD-1074; CI `python-runtime-compat` gate diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall index 2fa2b94..a12990e 100644 --- a/installer/macos/scripts/postinstall +++ b/installer/macos/scripts/postinstall @@ -17,13 +17,13 @@ AGENT_PATH="$AGENT_DIR/com.selectpilot.bridge.plist" HASH="$(shasum -a 256 "$SOURCE_ROOT/server/nano_server.py" | awk '{print $1}')" PYTHON_BIN="" for candidate in /opt/homebrew/bin/python3 /usr/local/bin/python3 /Library/Frameworks/Python.framework/Versions/Current/bin/python3 /usr/bin/python3; do - if [ -x "$candidate" ] && "$candidate" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, 10) else 1)'; then + if [ -x "$candidate" ] && "$candidate" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, 9) else 1)'; then PYTHON_BIN="$candidate" break fi done if [ -z "$PYTHON_BIN" ]; then - echo "SelectPilot requires Python 3.10 or newer." >&2 + echo "SelectPilot requires Python 3.9 or newer." >&2 exit 1 fi diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index d29e79f..4c1e110 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -25,8 +25,8 @@ RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}" PYTHON_BIN="${CHROMEAI_PYTHON_BIN:-$(command -v python3)}" -if ! "$PYTHON_BIN" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, 10) else 1)'; then - echo "SelectPilot requires Python 3.10 or newer; found $($PYTHON_BIN --version 2>&1)." >&2 +if ! "$PYTHON_BIN" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, 9) else 1)'; then + echo "SelectPilot requires Python 3.9 or newer; found $($PYTHON_BIN --version 2>&1)." >&2 exit 1 fi diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index 5adbe83..e20b3d4 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -67,7 +67,7 @@ test('macOS service installs outside the source checkout before launchd executio assert.match(installer, /presets\/extraction-presets\.json/); assert.match(installer, /model_policy\.json/); assert.match(installer, /model_registry\.runtime\.json/); - assert.match(installer, /sys\.version_info >= \(3, 10\)/); + assert.match(installer, /sys\.version_info >= \(3, 9\)/); assert.match(launchAgent, /__PYTHON_BIN__/); assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); assert.match(installer, /install -m 0555 "\$ROOT\/server\/\$module" "\$INSTALL_DIR\/\$module"/); @@ -76,11 +76,11 @@ test('macOS service installs outside the source checkout before launchd executio assert.doesNotMatch(launchAgent, /__PROJECT_ROOT__/); }); -test('production macOS package also binds launchd to Python 3.10 or newer', () => { +test('production macOS package also binds launchd to Python 3.9 or newer', () => { const postinstall = readFileSync('./installer/macos/scripts/postinstall', 'utf8'); const launchAgent = readFileSync('./installer/macos/com.selectpilot.bridge.plist.template', 'utf8'); - assert.match(postinstall, /sys\.version_info >= \(3, 10\)/); + assert.match(postinstall, /sys\.version_info >= \(3, 9\)/); assert.match(postinstall, /s\|__PYTHON_BIN__\|\$PYTHON_BIN\|g/); assert.match(launchAgent, /__PYTHON_BIN__/); assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); From 937e4926944052495c8fc626f2d36cb422fbec4d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Tue, 25 Aug 2026 22:33:39 +0200 Subject: [PATCH 47/66] feat(macos): add notarized helper release gate Keep CI candidate packaging unsigned while requiring an authorized Developer ID Installer identity and notarytool keychain profile for production release. Verified: shellcheck scripts/*.sh; node --test tests/panel/macos-release-contract.test.mjs tests/panel/bootstrap-contract.test.mjs; pnpm lint. --- .github/workflows/macos-helper.yml | 8 ++++- docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md | 15 +++++++- package.json | 1 + scripts/release-macos-helper.sh | 27 ++++++++++++++ tests/panel/macos-release-contract.test.mjs | 40 +++++++++++++++++++++ 5 files changed, 89 insertions(+), 2 deletions(-) create mode 100755 scripts/release-macos-helper.sh create mode 100644 tests/panel/macos-release-contract.test.mjs diff --git a/.github/workflows/macos-helper.yml b/.github/workflows/macos-helper.yml index 461ba7b..42e9b88 100644 --- a/.github/workflows/macos-helper.yml +++ b/.github/workflows/macos-helper.yml @@ -4,7 +4,7 @@ on: workflow_dispatch: pull_request: branches: [main] - paths: ["installer/macos/**", "server/**", "scripts/package-macos-helper.sh", ".github/workflows/macos-helper.yml"] + paths: ["installer/macos/**", "server/**", "scripts/package-macos-helper.sh", "scripts/release-macos-helper.sh", ".github/workflows/macos-helper.yml"] permissions: contents: read @@ -21,6 +21,12 @@ jobs: run: sh scripts/package-macos-helper.sh - name: Inspect package run: pkgutil --payload-files dist/macos-helper/SelectPilot-Installer-unsigned.pkg | sort + - name: Verify production release remains fail-closed without identities + run: | + if sh scripts/release-macos-helper.sh; then + echo "Production release unexpectedly succeeded without signing authority." >&2 + exit 1 + fi - uses: actions/upload-artifact@v7 with: name: selectpilot-macos-helper-unsigned-${{ github.sha }} diff --git a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md index b9b2f8a..f2e2ea6 100644 --- a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md +++ b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md @@ -1,6 +1,6 @@ # Chrome Web Store Release Checklist -Last updated: 2026-08-21 +Last updated: 2026-08-25 ## Automated Gate @@ -26,6 +26,19 @@ Never place the private key, its contents, or a production signature fixture in The separately operated local billing authority installs its pinned runtime with `python3 -m pip install --requirement requirements-billing.txt`. Its wallet RPC is loopback-only, entitlement/order state is written with owner-only permissions, administrative revocation requires a configured secret of at least 32 characters, and generated order IDs contain 128 bits of random material. +## macOS Helper Release Gate + +`pnpm package:helper:macos` creates the unsigned CI inspection artifact. It is not a production installer. + +The production release requires both of these values from authorized local signing state: + +- `SELECTPILOT_INSTALLER_SIGN_IDENTITY`: the exact Developer ID Installer identity; +- `SELECTPILOT_NOTARY_KEYCHAIN_PROFILE`: a `notarytool` keychain profile created outside the repository. + +Run `pnpm release:helper:macos` only on the signing Mac. The command fails before packaging when either authority is absent, signs the exact package, waits for Apple notarization, staples and validates the ticket, checks the package signature, runs Gatekeeper assessment, and prints the final SHA-256. Preserve that digest with the release evidence. + +Never upload the unsigned CI artifact or substitute an Apple Development identity. + ## Product Truth - [ ] The listing describes one purpose: selected text to structured local output. diff --git a/package.json b/package.json index 7e79b19..6995ee0 100644 --- a/package.json +++ b/package.json @@ -34,6 +34,7 @@ "entitlements:verify-key": "node ./scripts/verify-entitlement-key.mjs", "site:build": "node ./scripts/build-site-config.mjs", "package:helper:macos": "sh ./scripts/package-macos-helper.sh", + "release:helper:macos": "sh ./scripts/release-macos-helper.sh", "test:e2e": "playwright test --config tests/e2e/playwright.config.mjs", "test": "tsc -p tsconfig.json && node --test tests/panel/*.test.mjs && python3 -m unittest discover -s tests/server -p 'test_*.py'", "test:privacy": "node --test tests/panel/network-boundary.test.mjs && python3 -m unittest tests/server/test_privacy_proof.py", diff --git a/scripts/release-macos-helper.sh b/scripts/release-macos-helper.sh new file mode 100755 index 0000000..93f77b1 --- /dev/null +++ b/scripts/release-macos-helper.sh @@ -0,0 +1,27 @@ +#!/bin/sh +set -eu + +: "${SELECTPILOT_INSTALLER_SIGN_IDENTITY:?Set the exact Developer ID Installer identity.}" +: "${SELECTPILOT_NOTARY_KEYCHAIN_PROFILE:?Set an authorized notarytool keychain profile.}" + +ROOT="$(CDPATH='' cd -- "$(dirname "$0")/.." && pwd)" +SIGNED="$ROOT/dist/macos-helper/SelectPilot-Installer.pkg" + +SELECTPILOT_INSTALLER_SIGN_IDENTITY="$SELECTPILOT_INSTALLER_SIGN_IDENTITY" \ + sh "$ROOT/scripts/package-macos-helper.sh" + +if [ ! -f "$SIGNED" ]; then + echo "Signed SelectPilot installer was not produced." >&2 + exit 1 +fi + +xcrun notarytool submit "$SIGNED" \ + --keychain-profile "$SELECTPILOT_NOTARY_KEYCHAIN_PROFILE" \ + --wait +xcrun stapler staple "$SIGNED" +xcrun stapler validate "$SIGNED" +pkgutil --check-signature "$SIGNED" +spctl --assess --type install --verbose=2 "$SIGNED" + +DIGEST="$(shasum -a 256 "$SIGNED" | awk '{print $1}')" +printf '%s %s\n' "$SIGNED" "$DIGEST" diff --git a/tests/panel/macos-release-contract.test.mjs b/tests/panel/macos-release-contract.test.mjs new file mode 100644 index 0000000..a0ab65e --- /dev/null +++ b/tests/panel/macos-release-contract.test.mjs @@ -0,0 +1,40 @@ +// module_name: macos_release_contract_test +// spec_ref: "security" + +import assert from 'node:assert/strict'; +import { spawnSync } from 'node:child_process'; +import { readFileSync } from 'node:fs'; +import test from 'node:test'; + +test('production helper release fails before packaging without signing authority', () => { + const result = spawnSync('sh', ['./scripts/release-macos-helper.sh'], { + cwd: process.cwd(), + encoding: 'utf8', + env: { PATH: process.env.PATH ?? '' }, + }); + + assert.notEqual(result.status, 0); + assert.match(result.stderr, /SELECTPILOT_INSTALLER_SIGN_IDENTITY/); +}); + +test('production helper release requires notarization and Gatekeeper validation', () => { + const packageJson = JSON.parse(readFileSync('./package.json', 'utf8')); + const release = readFileSync('./scripts/release-macos-helper.sh', 'utf8'); + const candidate = readFileSync('./scripts/package-macos-helper.sh', 'utf8'); + + assert.equal(packageJson.scripts['release:helper:macos'], 'sh ./scripts/release-macos-helper.sh'); + assert.match(release, /SELECTPILOT_INSTALLER_SIGN_IDENTITY/); + assert.match(release, /SELECTPILOT_NOTARY_KEYCHAIN_PROFILE/); + assert.match(release, /package-macos-helper\.sh/); + assert.match(release, /notarytool submit/); + assert.match(release, /--keychain-profile/); + assert.match(release, /--wait/); + assert.match(release, /stapler staple/); + assert.match(release, /stapler validate/); + assert.match(release, /pkgutil --check-signature/); + assert.match(release, /spctl --assess --type install/); + assert.match(release, /shasum -a 256/); + + assert.match(candidate, /SelectPilot-Installer-unsigned\.pkg/); + assert.match(candidate, /SELECTPILOT_INSTALLER_SIGN_IDENTITY/); +}); From ed97789b8daec28425117274e3b06bf579ab6777 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 12:38:00 +0200 Subject: [PATCH 48/66] feat(release): add production entitlement deployment Add conflict-free loopback routing, hardened systemd and Nginx contracts, fail-closed preflight, safe diagnostics, and deployment regression coverage. Verified: pnpm lint; pnpm test; pnpm typecheck; pnpm test:privacy; pnpm package:store; pnpm test:e2e; shellcheck; systemd-analyze verify; hardened systemd smoke and recovery. --- docs/ENTITLEMENT_DEPLOYMENT.md | 45 ++++++++- package.json | 1 + scripts/install-entitlement-systemd.sh | 45 +++++++++ scripts/preflight-entitlement-deployment.sh | 91 +++++++++++++++++++ services/entitlement-authority/.env.example | 8 ++ services/entitlement-authority/Dockerfile | 3 +- services/entitlement-authority/compose.yaml | 10 +- .../nginx-http-bootstrap.conf | 13 +++ services/entitlement-authority/nginx-tls.conf | 22 +++++ .../selectpilot-entitlement.service | 51 +++++++++++ services/entitlement-authority/server.mjs | 10 +- tests/panel/production-deployment.test.mjs | 53 +++++++++++ 12 files changed, 347 insertions(+), 5 deletions(-) create mode 100755 scripts/install-entitlement-systemd.sh create mode 100755 scripts/preflight-entitlement-deployment.sh create mode 100644 services/entitlement-authority/.env.example create mode 100644 services/entitlement-authority/nginx-http-bootstrap.conf create mode 100644 services/entitlement-authority/nginx-tls.conf create mode 100644 services/entitlement-authority/selectpilot-entitlement.service create mode 100644 tests/panel/production-deployment.test.mjs diff --git a/docs/ENTITLEMENT_DEPLOYMENT.md b/docs/ENTITLEMENT_DEPLOYMENT.md index 620cb48..5f0ba03 100644 --- a/docs/ENTITLEMENT_DEPLOYMENT.md +++ b/docs/ENTITLEMENT_DEPLOYMENT.md @@ -19,7 +19,50 @@ The price map shape is: } ``` -Set the three corresponding `*_FILE` variables and `SELECTPILOT_STATE_DIR` in a local `.env`. Create the state directory with mode `0700` and ownership matching `SELECTPILOT_UID`/`SELECTPILOT_GID`, then run the Compose contract. Keep port `8090` bound to loopback and terminate public TLS in the existing reverse proxy at `license.selectpilot.app`. +Copy `services/entitlement-authority/.env.example` to an external deployment `.env` and set the three corresponding `*_FILE` variables. The file contains paths, IDs, and ports only; secret values remain in the referenced mode-`0600` files. Create the state directory with mode `0700` and ownership matching `SELECTPILOT_UID`/`SELECTPILOT_GID`. + +Run the fail-closed preflight before Compose: + +```sh +sh scripts/preflight-entitlement-deployment.sh /absolute/path/to/selectpilot-authority.env +docker compose --env-file /absolute/path/to/selectpilot-authority.env \ + -f services/entitlement-authority/compose.yaml up -d --build +``` + +The default host binding is `127.0.0.1:8091`; the container still listens on `8090`. Override `SELECTPILOT_AUTHORITY_HOST_PORT` only after confirming the replacement loopback port is free. Never bind the authority directly to a public interface. + +The production reverse proxy target is `http://127.0.0.1:8091`. Versioned Nginx bootstrap and TLS contracts are provided beside the Compose file. Install the HTTP bootstrap first, issue the certificate, then activate the TLS contract and run `nginx -t` before reload. + +### Hetzner production service + +The current Hetzner Snap-Docker daemon rejects executable startup whenever `no-new-privileges` is enabled. Disabling that control is not an accepted workaround. The production lane on this host is therefore the versioned native systemd service, which preserves `NoNewPrivileges`, loopback-only networking, read-only system paths, systemd credentials, an isolated dynamic user, and a private state directory. + +Place the three mode-`0600` production credential files at the exact paths declared by `selectpilot-entitlement.service`, then run: + +```sh +sudo sh scripts/install-entitlement-systemd.sh +``` + +The installer verifies the signing key against the pinned public identity, validates the hardened unit, starts it, and requires a successful loopback health request. It does not configure public DNS, TLS, Paddle, or Store publication. + +## DNS and TLS + +The authoritative DNS zone must contain these records before certificate issuance: + +| Name | Type | Value | +| --- | --- | --- | +| `license` | `A` | `46.62.191.13` | +| `license` | `AAAA` | `2a01:4f9:3090:1eab::2` | + +Verify both records resolve, then issue the certificate on Hetzner: + +```sh +sudo install -d -m 0755 /var/www/letsencrypt +sudo certbot certonly --webroot -w /var/www/letsencrypt -d license.selectpilot.app +sudo nginx -t +``` + +Do not request a certificate before public DNS resolves to the declared host. Required public routes: diff --git a/package.json b/package.json index 6995ee0..5b30f72 100644 --- a/package.json +++ b/package.json @@ -32,6 +32,7 @@ "package:store": "pnpm build && pnpm lint:manifest && node ./scripts/package-chrome-store.mjs", "entitlements:serve": "node ./services/entitlement-authority/server.mjs", "entitlements:verify-key": "node ./scripts/verify-entitlement-key.mjs", + "entitlements:preflight": "sh ./scripts/preflight-entitlement-deployment.sh", "site:build": "node ./scripts/build-site-config.mjs", "package:helper:macos": "sh ./scripts/package-macos-helper.sh", "release:helper:macos": "sh ./scripts/release-macos-helper.sh", diff --git a/scripts/install-entitlement-systemd.sh b/scripts/install-entitlement-systemd.sh new file mode 100755 index 0000000..8216564 --- /dev/null +++ b/scripts/install-entitlement-systemd.sh @@ -0,0 +1,45 @@ +#!/bin/sh +set -eu + +[ "$(id -u)" -eq 0 ] || { + echo "Run this installer as root." >&2 + exit 1 +} + +ROOT="$(CDPATH='' cd -- "$(dirname "$0")/.." && pwd)" +APP_DIR=/opt/selectpilot-entitlement +UNIT_SOURCE="$ROOT/services/entitlement-authority/selectpilot-entitlement.service" +UNIT_TARGET=/etc/systemd/system/selectpilot-entitlement.service + +for secret_file in \ + /etc/selectpilot/secrets/entitlement-ed25519.pem \ + /etc/selectpilot/secrets/paddle-webhook-secret \ + /etc/selectpilot/secrets/paddle-price-map.json +do + [ -s "$secret_file" ] || { + echo "Required production credential is missing." >&2 + exit 1 + } + [ "$(stat -c '%a' "$secret_file")" = "600" ] || { + echo "Production credentials must use mode 0600." >&2 + exit 1 + } +done + +if ss -ltn | awk '{print $4}' | grep -Eq '(^|:)8091$'; then + echo "Loopback port 8091 is already in use." >&2 + exit 1 +fi + +SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE=/etc/selectpilot/secrets/entitlement-ed25519.pem \ + node "$ROOT/scripts/verify-entitlement-key.mjs" + +install -d -o root -g root -m 0755 "$APP_DIR" +install -o root -g root -m 0644 "$ROOT/services/entitlement-authority/server.mjs" "$APP_DIR/server.mjs" +install -o root -g root -m 0644 "$UNIT_SOURCE" "$UNIT_TARGET" +systemd-analyze verify "$UNIT_TARGET" +systemctl daemon-reload +systemctl enable --now selectpilot-entitlement.service +systemctl is-active --quiet selectpilot-entitlement.service +curl --fail --silent --show-error http://127.0.0.1:8091/health >/dev/null +echo "SelectPilot entitlement authority is active on 127.0.0.1:8091." diff --git a/scripts/preflight-entitlement-deployment.sh b/scripts/preflight-entitlement-deployment.sh new file mode 100755 index 0000000..f25560d --- /dev/null +++ b/scripts/preflight-entitlement-deployment.sh @@ -0,0 +1,91 @@ +#!/bin/sh +set -eu + +ROOT="$(CDPATH='' cd -- "$(dirname "$0")/.." && pwd)" +COMPOSE_FILE="$ROOT/services/entitlement-authority/compose.yaml" +ENV_FILE="${1:-$ROOT/services/entitlement-authority/.env}" + +fail() { + printf 'Preflight failed: %s\n' "$1" >&2 + exit 1 +} + +[ -f "$ENV_FILE" ] || fail "missing environment file" +set -a +# The deployment environment file contains paths and scalar configuration only. +# shellcheck source=/dev/null +. "$ENV_FILE" +set +a + +for name in \ + SELECTPILOT_UID \ + SELECTPILOT_GID \ + SELECTPILOT_AUTHORITY_HOST_PORT \ + SELECTPILOT_ENTITLEMENT_KEY_ID \ + SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE \ + SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE \ + SELECTPILOT_PADDLE_PRICE_MAP_FILE \ + SELECTPILOT_STATE_DIR +do + eval "value=\${$name:-}" + [ -n "$value" ] || fail "missing $name" +done + +case "$SELECTPILOT_AUTHORITY_HOST_PORT" in + ''|*[!0-9]*) fail "authority host port must be numeric" ;; +esac +if [ "$SELECTPILOT_AUTHORITY_HOST_PORT" -lt 1024 ] || [ "$SELECTPILOT_AUTHORITY_HOST_PORT" -gt 65535 ]; then + fail "authority host port must be between 1024 and 65535" +fi + +file_mode() { + if [ "$(uname -s)" = "Darwin" ]; then + stat -f '%Lp' "$1" + else + stat -c '%a' "$1" + fi +} + +file_owner() { + if [ "$(uname -s)" = "Darwin" ]; then + stat -f '%u:%g' "$1" + else + stat -c '%u:%g' "$1" + fi +} + +for secret_file in \ + "$SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE" \ + "$SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE" \ + "$SELECTPILOT_PADDLE_PRICE_MAP_FILE" +do + [ -f "$secret_file" ] || fail "required secret file is missing" + [ -s "$secret_file" ] || fail "required secret file is empty" + [ "$(file_mode "$secret_file")" = "600" ] || fail "secret files must use mode 0600" +done + +[ -d "$SELECTPILOT_STATE_DIR" ] || fail "state directory is missing" +[ "$(file_mode "$SELECTPILOT_STATE_DIR")" = "700" ] || fail "state directory must use mode 0700" +[ "$(file_owner "$SELECTPILOT_STATE_DIR")" = "$SELECTPILOT_UID:$SELECTPILOT_GID" ] \ + || fail "state directory ownership does not match SELECTPILOT_UID:SELECTPILOT_GID" + +if command -v ss >/dev/null 2>&1 && ss -ltn | awk '{print $4}' | grep -Eq "(^|:)$SELECTPILOT_AUTHORITY_HOST_PORT$"; then + fail "authority host port is already in use" +fi + +node - "$SELECTPILOT_PADDLE_PRICE_MAP_FILE" <<'NODE' +import { readFileSync } from 'node:fs'; +const path = process.argv[2]; +const map = JSON.parse(readFileSync(path, 'utf8')); +if (!map.__trial__ || !['essential', 'plus', 'pro'].includes(map.__trial__.tier)) process.exit(1); +for (const [id, entry] of Object.entries(map)) { + if (id !== '__trial__' && !/^pri_[A-Za-z0-9]+$/.test(id)) process.exit(1); + if (!['essential', 'plus', 'pro'].includes(entry?.tier) || !Array.isArray(entry.features)) process.exit(1); +} +NODE + +SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE="$SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE" \ + node "$ROOT/scripts/verify-entitlement-key.mjs" + +docker compose --env-file "$ENV_FILE" -f "$COMPOSE_FILE" config --quiet +printf 'SelectPilot entitlement deployment preflight passed on loopback port %s.\n' "$SELECTPILOT_AUTHORITY_HOST_PORT" diff --git a/services/entitlement-authority/.env.example b/services/entitlement-authority/.env.example new file mode 100644 index 0000000..345ea23 --- /dev/null +++ b/services/entitlement-authority/.env.example @@ -0,0 +1,8 @@ +SELECTPILOT_UID=1000 +SELECTPILOT_GID=1000 +SELECTPILOT_AUTHORITY_HOST_PORT=8091 +SELECTPILOT_ENTITLEMENT_KEY_ID=selectpilot-prod-2026-01 +SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE=/etc/selectpilot/secrets/entitlement-ed25519.pem +SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE=/etc/selectpilot/secrets/paddle-webhook-secret +SELECTPILOT_PADDLE_PRICE_MAP_FILE=/etc/selectpilot/secrets/paddle-price-map.json +SELECTPILOT_STATE_DIR=/var/lib/selectpilot-authority diff --git a/services/entitlement-authority/Dockerfile b/services/entitlement-authority/Dockerfile index d6f18fc..944ba94 100644 --- a/services/entitlement-authority/Dockerfile +++ b/services/entitlement-authority/Dockerfile @@ -4,4 +4,5 @@ COPY services/entitlement-authority/server.mjs ./server.mjs USER node ENV HOST=0.0.0.0 PORT=8090 EXPOSE 8090 -CMD ["node", "server.mjs"] +ENTRYPOINT ["node"] +CMD ["server.mjs"] diff --git a/services/entitlement-authority/compose.yaml b/services/entitlement-authority/compose.yaml index 2629985..b20dc79 100644 --- a/services/entitlement-authority/compose.yaml +++ b/services/entitlement-authority/compose.yaml @@ -7,12 +7,12 @@ services: user: "${SELECTPILOT_UID:-1000}:${SELECTPILOT_GID:-1000}" read_only: true ports: - - "127.0.0.1:8090:8090" + - "127.0.0.1:${SELECTPILOT_AUTHORITY_HOST_PORT:-8091}:8090" environment: HOST: 0.0.0.0 PORT: "8090" SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE: /run/secrets/entitlement_private_key - SELECTPILOT_ENTITLEMENT_KEY_ID: selectpilot-prod-2026-01 + SELECTPILOT_ENTITLEMENT_KEY_ID: ${SELECTPILOT_ENTITLEMENT_KEY_ID:-selectpilot-prod-2026-01} SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE: /run/secrets/paddle_webhook_secret SELECTPILOT_PADDLE_PRICE_MAP_FILE: /run/secrets/paddle_price_map SELECTPILOT_ENTITLEMENT_STATE_FILE: /var/lib/selectpilot/entitlements.json @@ -29,6 +29,12 @@ services: - no-new-privileges:true cap_drop: - ALL + healthcheck: + test: ["CMD", "node", "-e", "fetch('http://127.0.0.1:8090/health').then(r=>{if(!r.ok)process.exit(1)}).catch(()=>process.exit(1))"] + interval: 15s + timeout: 3s + retries: 4 + start_period: 10s secrets: entitlement_private_key: diff --git a/services/entitlement-authority/nginx-http-bootstrap.conf b/services/entitlement-authority/nginx-http-bootstrap.conf new file mode 100644 index 0000000..144ee09 --- /dev/null +++ b/services/entitlement-authority/nginx-http-bootstrap.conf @@ -0,0 +1,13 @@ +server { + listen 46.62.191.13:80; + listen [2a01:4f9:3090:1eab::2]:80; + server_name license.selectpilot.app; + + location /.well-known/acme-challenge/ { + root /var/www/letsencrypt; + } + + location / { + return 308 https://license.selectpilot.app$request_uri; + } +} diff --git a/services/entitlement-authority/nginx-tls.conf b/services/entitlement-authority/nginx-tls.conf new file mode 100644 index 0000000..37fc76a --- /dev/null +++ b/services/entitlement-authority/nginx-tls.conf @@ -0,0 +1,22 @@ +server { + listen 46.62.191.13:443 ssl http2; + listen [2a01:4f9:3090:1eab::2]:443 ssl http2; + server_name license.selectpilot.app; + + ssl_certificate /etc/letsencrypt/live/license.selectpilot.app/fullchain.pem; + ssl_certificate_key /etc/letsencrypt/live/license.selectpilot.app/privkey.pem; + ssl_protocols TLSv1.2 TLSv1.3; + + client_max_body_size 256k; + + location / { + proxy_pass http://127.0.0.1:8091; + proxy_http_version 1.1; + proxy_set_header Host $host; + proxy_set_header X-Forwarded-Proto https; + proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for; + proxy_hide_header X-Powered-By; + add_header Strict-Transport-Security "max-age=31536000" always; + add_header X-Content-Type-Options nosniff always; + } +} diff --git a/services/entitlement-authority/selectpilot-entitlement.service b/services/entitlement-authority/selectpilot-entitlement.service new file mode 100644 index 0000000..271a51a --- /dev/null +++ b/services/entitlement-authority/selectpilot-entitlement.service @@ -0,0 +1,51 @@ +[Unit] +Description=SelectPilot entitlement authority +After=network-online.target +Wants=network-online.target + +[Service] +Type=simple +DynamicUser=yes +WorkingDirectory=/opt/selectpilot-entitlement +ExecStart=/usr/bin/node /opt/selectpilot-entitlement/server.mjs +Environment=HOST=127.0.0.1 +Environment=PORT=8091 +Environment=SELECTPILOT_ENTITLEMENT_KEY_ID=selectpilot-prod-2026-01 +Environment=SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE=%d/entitlement_private_key +Environment=SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE=%d/paddle_webhook_secret +Environment=SELECTPILOT_PADDLE_PRICE_MAP_FILE=%d/paddle_price_map +Environment=SELECTPILOT_ENTITLEMENT_STATE_FILE=/var/lib/selectpilot-authority/entitlements.json +Environment=SELECTPILOT_WEB_ORIGIN=https://selectpilot.app +LoadCredential=entitlement_private_key:/etc/selectpilot/secrets/entitlement-ed25519.pem +LoadCredential=paddle_webhook_secret:/etc/selectpilot/secrets/paddle-webhook-secret +LoadCredential=paddle_price_map:/etc/selectpilot/secrets/paddle-price-map.json +StateDirectory=selectpilot-authority +StateDirectoryMode=0700 +UMask=0077 +NoNewPrivileges=yes +PrivateDevices=yes +PrivateTmp=yes +ProtectClock=yes +ProtectControlGroups=yes +ProtectHome=yes +ProtectHostname=yes +ProtectKernelLogs=yes +ProtectKernelModules=yes +ProtectKernelTunables=yes +ProtectProc=invisible +ProtectSystem=strict +RestrictAddressFamilies=AF_INET AF_INET6 +RestrictNamespaces=yes +RestrictRealtime=yes +RestrictSUIDSGID=yes +CapabilityBoundingSet= +LockPersonality=yes +SystemCallArchitectures=native +SystemCallFilter=@system-service +IPAddressDeny=any +IPAddressAllow=localhost +Restart=on-failure +RestartSec=3s + +[Install] +WantedBy=multi-user.target diff --git a/services/entitlement-authority/server.mjs b/services/entitlement-authority/server.mjs index 494d193..eb43ff4 100644 --- a/services/entitlement-authority/server.mjs +++ b/services/entitlement-authority/server.mjs @@ -277,7 +277,15 @@ export async function createAuthority(config = {}) { return send(res, 404, { error: 'not_found' }); } catch (error) { - send(res, error.status || 500, { error: error.status ? error.message : 'internal_error' }); + const status = error.status || 500; + console.error(JSON.stringify({ + event: 'entitlement_authority_request_failed', + method: req.method, + path: req.url, + status, + error: error.code || error.name || 'Error', + })); + send(res, status, { error: error.status ? error.message : 'internal_error' }); } }); } diff --git a/tests/panel/production-deployment.test.mjs b/tests/panel/production-deployment.test.mjs new file mode 100644 index 0000000..4e38767 --- /dev/null +++ b/tests/panel/production-deployment.test.mjs @@ -0,0 +1,53 @@ +// module_name: production_deployment_contract +// spec_ref: "validation_layer" +import assert from 'node:assert/strict'; +import { readFile, stat } from 'node:fs/promises'; +import test from 'node:test'; + +test('entitlement authority uses a configurable conflict-free loopback port', async () => { + const compose = await readFile('services/entitlement-authority/compose.yaml', 'utf8'); + const dockerfile = await readFile('services/entitlement-authority/Dockerfile', 'utf8'); + assert.match(compose, /127\.0\.0\.1:\$\{SELECTPILOT_AUTHORITY_HOST_PORT:-8091\}:8090/); + assert.match(compose, /healthcheck:/); + assert.doesNotMatch(compose, /127\.0\.0\.1:8090:8090/); + assert.match(dockerfile, /ENTRYPOINT \["node"\]/); + assert.doesNotMatch(dockerfile, /CMD \["node", "server\.mjs"\]/); +}); + +test('production proxy contract preserves TLS and loopback boundaries', async () => { + const nginx = await readFile('services/entitlement-authority/nginx-tls.conf', 'utf8'); + assert.match(nginx, /server_name license\.selectpilot\.app/); + assert.match(nginx, /proxy_pass http:\/\/127\.0\.0\.1:8091/); + assert.match(nginx, /ssl_protocols TLSv1\.2 TLSv1\.3/); + assert.doesNotMatch(nginx, /proxy_pass http:\/\/(?!127\.0\.0\.1)/); +}); + +test('Hetzner systemd authority preserves the security boundary', async () => { + const unit = await readFile('services/entitlement-authority/selectpilot-entitlement.service', 'utf8'); + assert.match(unit, /DynamicUser=yes/); + assert.match(unit, /NoNewPrivileges=yes/); + assert.match(unit, /ProtectSystem=strict/); + assert.match(unit, /IPAddressDeny=any/); + assert.match(unit, /IPAddressAllow=localhost/); + assert.match(unit, /LoadCredential=entitlement_private_key:/); + assert.match(unit, /Environment=HOST=127\.0\.0\.1/); + assert.match(unit, /Environment=PORT=8091/); +}); + +test('authority diagnostics never log request or credential content', async () => { + const server = await readFile('services/entitlement-authority/server.mjs', 'utf8'); + assert.match(server, /entitlement_authority_request_failed/); + assert.match(server, /method: req\.method/); + assert.match(server, /path: req\.url/); + assert.doesNotMatch(server, /body: rawBody|token: token|privateKey:/); +}); + +test('deployment preflight is executable and never prints secret contents', async () => { + for (const scriptPath of ['scripts/preflight-entitlement-deployment.sh', 'scripts/install-entitlement-systemd.sh']) { + const script = await readFile(scriptPath, 'utf8'); + const metadata = await stat(scriptPath); + assert.ok(metadata.mode & 0o100); + assert.match(script, /verify-entitlement-key\.mjs/); + assert.doesNotMatch(script, /cat [^\n]*SECRET|set -x/); + } +}); From c2c7394f92d79199ac1cb62e3cfd9f5d9af86d6f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 12:44:09 +0200 Subject: [PATCH 49/66] fix(billing): harden Paddle webhook verification --- services/entitlement-authority/server.mjs | 14 ++++++++------ tests/panel/entitlement-authority.test.mjs | 11 +++++++++-- 2 files changed, 17 insertions(+), 8 deletions(-) diff --git a/services/entitlement-authority/server.mjs b/services/entitlement-authority/server.mjs index eb43ff4..c546006 100644 --- a/services/entitlement-authority/server.mjs +++ b/services/entitlement-authority/server.mjs @@ -8,7 +8,7 @@ import { mkdir } from 'node:fs/promises'; const TIERS = new Set(['essential', 'plus', 'pro']); const MAX_BODY_BYTES = 256 * 1024; -const PADDLE_TOLERANCE_SECONDS = 300; +const PADDLE_TOLERANCE_SECONDS = 5; function required(name, env = process.env) { const value = env[name]?.trim(); @@ -46,13 +46,15 @@ function parsePriceMap(value) { // @spec_ref validation_layer export function verifyPaddleSignature(rawBody, header, secret, nowSeconds = Math.floor(Date.now() / 1000)) { - const parts = Object.fromEntries(String(header || '').split(';').map((part) => part.split('=', 2))); - const timestamp = Number(parts.ts); - const supplied = parts.h1; - if (!Number.isInteger(timestamp) || !/^[0-9a-f]{64}$/i.test(supplied || '')) return false; + const parts = String(header || '').split(';').map((part) => part.split('=', 2)); + const timestamps = parts.filter(([name]) => name === 'ts').map(([, value]) => value); + const signatures = parts.filter(([name]) => name === 'h1').map(([, value]) => value); + const timestamp = Number(timestamps[0]); + if (timestamps.length !== 1 || signatures.length === 0 || !Number.isInteger(timestamp)) return false; + if (signatures.some((supplied) => !/^[0-9a-f]{64}$/i.test(supplied || ''))) return false; if (Math.abs(nowSeconds - timestamp) > PADDLE_TOLERANCE_SECONDS) return false; const expected = createHmac('sha256', secret).update(`${timestamp}:${rawBody}`).digest('hex'); - return timingSafeEqual(Buffer.from(expected, 'hex'), Buffer.from(supplied, 'hex')); + return signatures.some((supplied) => timingSafeEqual(Buffer.from(expected, 'hex'), Buffer.from(supplied, 'hex'))); } async function readBody(req) { diff --git a/tests/panel/entitlement-authority.test.mjs b/tests/panel/entitlement-authority.test.mjs index 50bb4a3..9dc6ea9 100644 --- a/tests/panel/entitlement-authority.test.mjs +++ b/tests/panel/entitlement-authority.test.mjs @@ -117,11 +117,18 @@ test('subscription updates preserve the token and cancellation revokes it', asyn assert.equal((await post(base, '/v1/entitlements/verify', { token: claimed.token })).status, 401); }); -test('Paddle signature rejects stale and forged events', () => { +test('Paddle signature supports rotation and rejects stale or forged events', () => { const body = '{}'; const now = 2_000_000_000; const valid = createHmac('sha256', 'secret').update(`${now}:${body}`).digest('hex'); + const invalid = createHmac('sha256', 'wrong').update(`${now}:${body}`).digest('hex'); + const boundary = createHmac('sha256', 'secret').update(`${now - 5}:${body}`).digest('hex'); + const stale = createHmac('sha256', 'secret').update(`${now - 6}:${body}`).digest('hex'); assert.equal(verifyPaddleSignature(body, `ts=${now};h1=${valid}`, 'secret', now), true); + assert.equal(verifyPaddleSignature(body, `ts=${now};h1=${invalid};h1=${valid}`, 'secret', now), true); assert.equal(verifyPaddleSignature(body, `ts=${now};h1=${valid}`, 'wrong', now), false); - assert.equal(verifyPaddleSignature(body, `ts=${now - 301};h1=${valid}`, 'secret', now), false); + assert.equal(verifyPaddleSignature(body, `ts=${now - 5};h1=${boundary}`, 'secret', now), true); + assert.equal(verifyPaddleSignature(body, `ts=${now - 6};h1=${stale}`, 'secret', now), false); + assert.equal(verifyPaddleSignature(body, `ts=${now};ts=${now};h1=${valid}`, 'secret', now), false); + assert.equal(verifyPaddleSignature(body, `ts=${now};h1=malformed;h1=${valid}`, 'secret', now), false); }); From 9856bd29af92df35c97f78f425b8edc99462c0f1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 12:44:52 +0200 Subject: [PATCH 50/66] fix(security): harden local privacy boundaries Intent: AXIOM-20260830-OXBH25 Encrypt project memory at rest, keep preferences device-local, redact runtime errors, bound bridge requests, and activate downstream Axiom enforcement. Not run: dependencies are intentionally absent on the Mac control surface; full verification runs on Hetzner. --- .github/workflows/axiom.yml | 18 +++++ api/request.js | 33 +++++++-- api/request.ts | 33 +++++++-- background/background.js | 15 +++- background/background.ts | 14 +++- .../10-intent.md | 40 +++++++++++ .../20-checkpoint.md | 22 ++++++ intent.json | 72 +++++++++++++++++++ options/options.js | 4 +- options/options.ts | 4 +- tests/panel/api-request-hardening.test.mjs | 35 +++++++++ tests/panel/local-settings-boundary.test.mjs | 10 +++ tests/panel/storage-encryption.test.mjs | 17 ++++- utils/logger.js | 23 ++++-- utils/logger.ts | 23 ++++-- utils/storage.js | 15 ++++ utils/storage.ts | 15 ++++ 17 files changed, 358 insertions(+), 35 deletions(-) create mode 100644 .github/workflows/axiom.yml create mode 100644 docs/aegis/work/2026-08-30-production-abuse-gate/10-intent.md create mode 100644 docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md create mode 100644 intent.json create mode 100644 tests/panel/api-request-hardening.test.mjs create mode 100644 tests/panel/local-settings-boundary.test.mjs diff --git a/.github/workflows/axiom.yml b/.github/workflows/axiom.yml new file mode 100644 index 0000000..cfc26c4 --- /dev/null +++ b/.github/workflows/axiom.yml @@ -0,0 +1,18 @@ +name: Axiom Enforcement + +on: + pull_request: + push: + branches: + - main + +permissions: + contents: read + pull-requests: read + +jobs: + axiom: + uses: kristoffersodersten/axiom-core/.github/workflows/axiom-downstream.yml@main + with: + intent_path: intent.json + strict: true diff --git a/api/request.js b/api/request.js index b872da0..bb2620b 100644 --- a/api/request.js +++ b/api/request.js @@ -1,6 +1,7 @@ // module_name: api_request_ts // spec_ref: "execution_layer" import { log, warn } from '../utils/logger.js'; +const DEFAULT_TIMEOUT_MS = 15_000; export class ApiRequestError extends Error { status; code; @@ -33,8 +34,13 @@ function parseJsonSafe(text) { } } export async function apiRequest(url, options = {}) { - const { method = 'POST', headers = {}, body } = options; + const { method = 'POST', headers = {}, body, timeoutMs = DEFAULT_TIMEOUT_MS } = options; + if (!Number.isFinite(timeoutMs) || timeoutMs < 100 || timeoutMs > 120_000) { + throw new TypeError('timeoutMs must be between 100 and 120000'); + } const traceId = createTraceId(); + const controller = new AbortController(); + const timeout = setTimeout(() => controller.abort(), timeoutMs); const init = { method, headers: { @@ -43,10 +49,25 @@ export async function apiRequest(url, options = {}) { ...headers }, body: body !== undefined ? JSON.stringify(body) : undefined, - cache: 'no-store' + cache: 'no-store', + signal: controller.signal, }; - log('api', method, url); - const res = await fetch(url, init); + log('api', method); + let res; + try { + res = await fetch(url, init); + } + catch (cause) { + const timedOut = cause instanceof DOMException && cause.name === 'AbortError'; + throw new ApiRequestError(timedOut ? 'Local runtime request timed out' : 'Local runtime request failed', { + status: 0, + code: timedOut ? 'request_timeout' : 'network_error', + traceId, + }); + } + finally { + clearTimeout(timeout); + } const text = await res.text(); const parsed = parseJsonSafe(text); const maybeError = parsed?.error; @@ -54,9 +75,9 @@ export async function apiRequest(url, options = {}) { maybeError?.details?.trace_id || traceId; if (!res.ok) { - warn('api', 'non-200', res.status, text); + warn('api', 'non_200', res.status); const code = maybeError?.code || `http_${res.status}`; - const message = maybeError?.message || `API ${res.status}: ${text}`; + const message = maybeError?.message || `Local runtime returned HTTP ${res.status}`; throw new ApiRequestError(message, { status: res.status, code, diff --git a/api/request.ts b/api/request.ts index 4aae3f6..36fe77b 100644 --- a/api/request.ts +++ b/api/request.ts @@ -6,8 +6,11 @@ export type RequestOptions = { method?: 'GET' | 'POST'; headers?: Record; body?: unknown; + timeoutMs?: number; }; +const DEFAULT_TIMEOUT_MS = 15_000; + type ApiErrorShape = { error?: { code?: string; @@ -54,8 +57,13 @@ function parseJsonSafe(text: string): Record | null { } export async function apiRequest(url: string, options: RequestOptions = {}): Promise { - const { method = 'POST', headers = {}, body } = options; + const { method = 'POST', headers = {}, body, timeoutMs = DEFAULT_TIMEOUT_MS } = options; + if (!Number.isFinite(timeoutMs) || timeoutMs < 100 || timeoutMs > 120_000) { + throw new TypeError('timeoutMs must be between 100 and 120000'); + } const traceId = createTraceId(); + const controller = new AbortController(); + const timeout = setTimeout(() => controller.abort(), timeoutMs); const init: RequestInit = { method, headers: { @@ -64,10 +72,23 @@ export async function apiRequest(url: string, options: RequestOptions = {}): ...headers }, body: body !== undefined ? JSON.stringify(body) : undefined, - cache: 'no-store' + cache: 'no-store', + signal: controller.signal, }; - log('api', method, url); - const res = await fetch(url, init); + log('api', method); + let res: Response; + try { + res = await fetch(url, init); + } catch (cause) { + const timedOut = cause instanceof DOMException && cause.name === 'AbortError'; + throw new ApiRequestError(timedOut ? 'Local runtime request timed out' : 'Local runtime request failed', { + status: 0, + code: timedOut ? 'request_timeout' : 'network_error', + traceId, + }); + } finally { + clearTimeout(timeout); + } const text = await res.text(); const parsed = parseJsonSafe(text); const maybeError = (parsed as ApiErrorShape | null)?.error; @@ -77,9 +98,9 @@ export async function apiRequest(url: string, options: RequestOptions = {}): traceId; if (!res.ok) { - warn('api', 'non-200', res.status, text); + warn('api', 'non_200', res.status); const code = maybeError?.code || `http_${res.status}`; - const message = maybeError?.message || `API ${res.status}: ${text}`; + const message = maybeError?.message || `Local runtime returned HTTP ${res.status}`; throw new ApiRequestError(message, { status: res.status, code, diff --git a/background/background.js b/background/background.js index 7a18f71..3582069 100644 --- a/background/background.js +++ b/background/background.js @@ -7,6 +7,7 @@ import { requireFeature, getLicenseTier, attachLicenseToken, refreshLicense } fr import { getEntitlementSnapshot } from './entitlement-service.js'; import { ApiRequestError } from '../api/request.js'; import { FIRST_RUN_EXAMPLE } from '../shared/first-run-example.js'; +import { getEncryptedJSON, setEncryptedJSON } from '../utils/storage.js'; const MEMORY_ENABLED_KEY = 'selectpilot_memory_enabled_v1'; const MEMORY_LEDGER_KEY = 'selectpilot_memory_ledger_v1'; async function openForActiveTab() { @@ -36,12 +37,20 @@ async function setMemoryEnabled(enabled) { await chrome.storage.local.set({ [MEMORY_ENABLED_KEY]: enabled }); } async function getMemoryLedger() { + const encrypted = await getEncryptedJSON(MEMORY_LEDGER_KEY); + if (Array.isArray(encrypted)) + return encrypted; + // One-way, in-place migration from the pre-encryption schema. const stored = await chrome.storage.local.get(MEMORY_LEDGER_KEY); - const entries = stored[MEMORY_LEDGER_KEY]; - return Array.isArray(entries) ? entries : []; + const legacy = stored[MEMORY_LEDGER_KEY]; + if (!Array.isArray(legacy)) + return []; + const bounded = legacy.slice(-200); + await setEncryptedJSON(MEMORY_LEDGER_KEY, bounded); + return bounded; } async function setMemoryLedger(entries) { - await chrome.storage.local.set({ [MEMORY_LEDGER_KEY]: entries.slice(-200) }); + await setEncryptedJSON(MEMORY_LEDGER_KEY, entries.slice(-200)); } async function recordMemoryEvent(entry) { if (!(await canUseProjectMemory())) diff --git a/background/background.ts b/background/background.ts index d5c2154..e57bbd1 100644 --- a/background/background.ts +++ b/background/background.ts @@ -8,6 +8,7 @@ import { getEntitlementSnapshot } from './entitlement-service.js'; import { ApiRequestError } from '../api/request.js'; import type { AgentContext } from '../agent/agent-types.js'; import { FIRST_RUN_EXAMPLE } from '../shared/first-run-example.js'; +import { getEncryptedJSON, setEncryptedJSON } from '../utils/storage.js'; type MemoryEntry = { action: 'extract' | 'summarize' | 'agent'; @@ -57,13 +58,20 @@ async function setMemoryEnabled(enabled: boolean): Promise { } async function getMemoryLedger(): Promise { + const encrypted = await getEncryptedJSON(MEMORY_LEDGER_KEY); + if (Array.isArray(encrypted)) return encrypted; + + // One-way, in-place migration from the pre-encryption schema. const stored = await chrome.storage.local.get(MEMORY_LEDGER_KEY); - const entries = stored[MEMORY_LEDGER_KEY]; - return Array.isArray(entries) ? entries : []; + const legacy = stored[MEMORY_LEDGER_KEY]; + if (!Array.isArray(legacy)) return []; + const bounded = legacy.slice(-200) as MemoryEntry[]; + await setEncryptedJSON(MEMORY_LEDGER_KEY, bounded); + return bounded; } async function setMemoryLedger(entries: MemoryEntry[]): Promise { - await chrome.storage.local.set({ [MEMORY_LEDGER_KEY]: entries.slice(-200) }); + await setEncryptedJSON(MEMORY_LEDGER_KEY, entries.slice(-200)); } async function recordMemoryEvent(entry: Omit): Promise { diff --git a/docs/aegis/work/2026-08-30-production-abuse-gate/10-intent.md b/docs/aegis/work/2026-08-30-production-abuse-gate/10-intent.md new file mode 100644 index 0000000..2794276 --- /dev/null +++ b/docs/aegis/work/2026-08-30-production-abuse-gate/10-intent.md @@ -0,0 +1,40 @@ +# SelectPilot Production Abuse Gate + +## TaskIntentDraft + +- Requested outcome: close all material global-governance, security, reliability, abuse, failure, recovery, and negative-space gaps before Chrome Web Store identity verification and publication continue. +- Scope: the complete SelectPilot product and release path represented by PR #22, including extension, local bridge, entitlement authority, installer, site, CI, packaging, runtime policy, evidence, and deployment contracts. +- Non-goals: no self-approval, branch-protection bypass, fabricated production credentials, silent fallback, production publication, or destructive live-state mutation. +- Signed intent: `AXIOM-20260830-OXBH25` in `intent.json`. +- Stop states: `done`, `blocked`, `needs-verification`, or `scope-exceeded`; only full-system evidence may support `done`. + +## BaselineReadSetHint + +- `/Users/kristoffersodersten/.codex/constitution-policy.yaml` +- global and repository `AGENTS.md` +- `CONTEXT.md` +- `docs/aegis/plans/2026-08-18-selectpilot-full-system-completion.md` +- `docs/selectpilot-project-snapshot.json` (known stale; evidence only) +- Linear project `SelectPilot — Lokal modell & runtime-stack` and SOD-1071 +- PR #22 exact state and current head +- signed Axiom NovaForge lifecycle policy + +## BaselineUsageDraft + +- Acknowledged: constitution, global/repository rules, parent plan, current PR state, current project context, signed lifecycle policy. +- Missing or drifted: installed NovaForge operator policy binding; project snapshot exact SHA/current state. +- Decision: fail closed on unsupported completion claims; continue read-only audit and signed-intent-authorized repository repair. + +## ImpactStatementDraft + +This work may change public release contracts, CI gates, failure handling, abuse resistance, installer/runtime behavior, and evidence generation. All heavy proof must execute on Hetzner or GitHub Actions; local work is limited to control, editing, signing, and bounded inspection. + +## Execution Readiness View + +- Intent lock: full-system production resilience before Store continuation. +- Scope fence: SelectPilot repository and its declared deployment/runtime contracts. +- Canonical owners: existing extension, bridge, entitlement, installer, site, policy, and CI modules; no parallel fallback engines. +- Compatibility boundary: published local bridge/extension and billing contracts remain explicit and fail closed. +- Test obligations: deterministic abuse/fault matrix, security/privacy, resource/concurrency, recovery/rollback, full regression, exact-head CI, and physical-runtime verification. +- Review gate: protected independent approval remains mandatory. +- Drift rule: any conflict among constitution, Linear, code, runtime, and evidence pauses the affected claim and repairs the authoritative owner. diff --git a/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md new file mode 100644 index 0000000..bfc4264 --- /dev/null +++ b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md @@ -0,0 +1,22 @@ +# Checkpoint + +## TodoCheckpointDraft + +- Current todo: inventory full architecture, contracts, tests, threat model, abuse surfaces, and negative space at exact head `937e4926944052495c8fc626f2d36cb422fbec4d`. +- Completed: project identity verified; branch fast-forwarded; global constitution read; signed lifecycle policy cryptographically verified; signed intent created; Store flow paused. +- Active slice: read-only gap and abuse-surface audit. +- Evidence: PR #22 remains open and review-blocked; signed Axiom intent `AXIOM-20260830-OXBH25`; lifecycle policy hash `932bbb231cb5e1aeb35d6bbe6327db6f3b1b3cf1e125e38366cbccbbc59434cf`. +- Blocked-on: installed NovaForge operator release references a missing `config/axiom-lifecycle-policy-binding.json`; production authorities remain external but do not block repository hardening. +- Next: enumerate existing controls and tests, run fresh remote baseline, then classify and repair material findings. + +## ResumeStateHint + +Resume from this checkpoint, re-read `10-intent.md`, verify `intent.json`, compare current head/worktree with this file, and refresh PR/Linear state before any completion claim. + +## DriftCheckDraft + +- Intent alignment: yes. +- Scope alignment: yes. +- New fallback or duplicate owner: no. +- Evidence sufficiency: not yet; audit and stress proof pending. +- Decision: continue. diff --git a/intent.json b/intent.json new file mode 100644 index 0000000..d1211e5 --- /dev/null +++ b/intent.json @@ -0,0 +1,72 @@ +{ + "id": "AXIOM-20260830-OXBH25", + "goal": "Close SelectPilot pre-deployment gaps, blind spots, negative space, abuse resilience, recovery, security, and full-system production readiness before Chrome Web Store publication", + "linear_issue_id": "SOD-1071", + "canonical_reference": "kristoffersodersten/selectpilot@937e4926944052495c8fc626f2d36cb422fbec4d", + "contract_target": "docs/aegis/plans/2026-08-18-selectpilot-full-system-completion.md", + "layer": "full-system-release-hardening", + "acceptance_criteria": [ + "global constitution mapped and enforced", + "material gaps implemented", + "extreme deterministic abuse suite passes", + "no implicit fallback or bypass", + "rollback and recovery proven", + "exact evidence published" + ], + "verification_method": [ + "Hetzner full test and stress suite", + "GitHub exact-head CI and security", + "physical Apple runtime failure and recovery", + "Linear SOD-1071 evidence" + ], + "constraints": [ + "no self approval", + "no branch protection bypass", + "no production claim before exact merge SHA", + "no secrets or PII in repo or logs", + "Mac control only and heavy proof on Hetzner or GitHub" + ], + "allowed_actions": [ + "**" + ], + "forbidden_actions": [ + ".env*", + "**/*.pem", + "**/*.key", + "**/secrets/**", + "main" + ], + "perception_targets": [ + "explicit causal errors", + "visible locality and provider", + "visible recovery state", + "calm non-coercive UI under failure" + ], + "expires_at": 1790726400, + "nonce": "a43cdb50-45fa-4d70-8cb4-241544a20873", + "signer": { + "public_key": "MCowBQYDK2VwAyEAqVq6w6JT3xe8AZydV7bto90lPToV8GqmeCmxG09zlgU=", + "role": "owner" + }, + "signature": { + "algorithm": "ed25519", + "public_key": "MCowBQYDK2VwAyEAqVq6w6JT3xe8AZydV7bto90lPToV8GqmeCmxG09zlgU=", + "value": "/La7J8OL82qBNOUXwQn05xub3QpM0GRVVbB7H7R2bVg6uKRr8J6rYLn0WkeuhrlX+PV++PC5Q0nmwxpNfuzgBA==", + "signed_payload_fields": [ + "id", + "goal", + "linear_issue_id", + "canonical_reference", + "contract_target", + "layer", + "acceptance_criteria", + "verification_method", + "constraints", + "allowed_actions", + "forbidden_actions", + "perception_targets", + "expires_at", + "nonce" + ] + } +} diff --git a/options/options.js b/options/options.js index 791c83e..43aab4e 100644 --- a/options/options.js +++ b/options/options.js @@ -8,7 +8,7 @@ const fields = { }; const statusMessage = document.querySelector('#status'); async function load() { - const stored = await chrome.storage.sync.get('selectpilot_settings'); + const stored = await chrome.storage.local.get('selectpilot_settings'); const settings = { ...defaults, ...(stored.selectpilot_settings || {}) }; fields.controlSide.value = settings.controlSide; fields.resultView.value = settings.resultView; @@ -24,7 +24,7 @@ async function save() { highContrast: fields.highContrast.checked, textSize: fields.textSize.value, }; - await chrome.storage.sync.set({ selectpilot_settings: settings }); + await chrome.storage.local.set({ selectpilot_settings: settings }); if (statusMessage) statusMessage.textContent = 'Saved.'; } diff --git a/options/options.ts b/options/options.ts index 5622be4..cb0b2c4 100644 --- a/options/options.ts +++ b/options/options.ts @@ -21,7 +21,7 @@ const fields = { const statusMessage = document.querySelector('#status'); async function load() { - const stored = await chrome.storage.sync.get('selectpilot_settings'); + const stored = await chrome.storage.local.get('selectpilot_settings'); const settings = { ...defaults, ...(stored.selectpilot_settings || {}) } as Settings; fields.controlSide!.value = settings.controlSide; fields.resultView!.value = settings.resultView; @@ -38,7 +38,7 @@ async function save() { highContrast: fields.highContrast!.checked, textSize: fields.textSize!.value as Settings['textSize'], }; - await chrome.storage.sync.set({ selectpilot_settings: settings }); + await chrome.storage.local.set({ selectpilot_settings: settings }); if (statusMessage) statusMessage.textContent = 'Saved.'; } diff --git a/tests/panel/api-request-hardening.test.mjs b/tests/panel/api-request-hardening.test.mjs new file mode 100644 index 0000000..30b3acc --- /dev/null +++ b/tests/panel/api-request-hardening.test.mjs @@ -0,0 +1,35 @@ +import assert from 'node:assert/strict'; +import test from 'node:test'; + +const { ApiRequestError, apiRequest } = await import('../../api/request.js'); + +test('request timeout aborts a hung local runtime explicitly', async (t) => { + const originalFetch = globalThis.fetch; + globalThis.fetch = (_url, init) => new Promise((_resolve, reject) => { + init.signal.addEventListener('abort', () => reject(new DOMException('aborted', 'AbortError'))); + }); + t.after(() => { globalThis.fetch = originalFetch; }); + + await assert.rejects( + apiRequest('http://127.0.0.1:11435/hung', { timeoutMs: 100 }), + (error) => error instanceof ApiRequestError && error.code === 'request_timeout' && error.status === 0, + ); +}); + +test('untrusted response bodies are not reflected in fallback errors', async (t) => { + const originalFetch = globalThis.fetch; + globalThis.fetch = async () => new Response('secret model output', { status: 502 }); + t.after(() => { globalThis.fetch = originalFetch; }); + + await assert.rejects( + apiRequest('http://127.0.0.1:11435/failure'), + (error) => error instanceof ApiRequestError + && error.code === 'http_502' + && !error.message.includes('secret model output'), + ); +}); + +test('timeout budget is bounded by contract', async () => { + await assert.rejects(apiRequest('http://127.0.0.1:11435', { timeoutMs: 0 }), TypeError); + await assert.rejects(apiRequest('http://127.0.0.1:11435', { timeoutMs: 120_001 }), TypeError); +}); diff --git a/tests/panel/local-settings-boundary.test.mjs b/tests/panel/local-settings-boundary.test.mjs new file mode 100644 index 0000000..cebd660 --- /dev/null +++ b/tests/panel/local-settings-boundary.test.mjs @@ -0,0 +1,10 @@ +import assert from 'node:assert/strict'; +import { readFile } from 'node:fs/promises'; +import test from 'node:test'; + +test('extension preferences remain device-local and never use Chrome sync', async () => { + const source = await readFile(new URL('../../options/options.ts', import.meta.url), 'utf8'); + assert.match(source, /chrome\.storage\.local\.get/); + assert.match(source, /chrome\.storage\.local\.set/); + assert.doesNotMatch(source, /chrome\.storage\.sync/); +}); diff --git a/tests/panel/storage-encryption.test.mjs b/tests/panel/storage-encryption.test.mjs index 6d34765..47eae00 100644 --- a/tests/panel/storage-encryption.test.mjs +++ b/tests/panel/storage-encryption.test.mjs @@ -17,7 +17,7 @@ globalThis.chrome = { }, }; -const { getDecrypted, setEncrypted } = await import('../../utils/storage.js'); +const { getDecrypted, getEncryptedJSON, setEncrypted, setEncryptedJSON } = await import('../../utils/storage.js'); test('AES-GCM storage creates a non-exportable per-install key and round-trips data', async () => { await setEncrypted('test_payload', 'private local value'); @@ -28,3 +28,18 @@ test('AES-GCM storage creates a non-exportable per-install key and round-trips d assert.equal(typeof values.get('test_payload'), 'string'); assert.equal(values.get('test_payload').includes('private local value'), false); }); + +test('encrypted JSON never persists private fields as plaintext', async () => { + const ledger = [{ content: 'private selected text', url: 'https://private.example/path' }]; + await setEncryptedJSON('ledger', ledger); + + const persisted = values.get('ledger'); + assert.equal(persisted.includes('private selected text'), false); + assert.equal(persisted.includes('private.example'), false); + assert.deepEqual(await getEncryptedJSON('ledger'), ledger); +}); + +test('corrupt encrypted JSON fails closed without returning partial data', async () => { + values.set('corrupt', JSON.stringify({ iv: [1, 2], data: [3, 4] })); + assert.equal(await getEncryptedJSON('corrupt'), null); +}); diff --git a/utils/logger.js b/utils/logger.js index deaed7f..bd4538f 100644 --- a/utils/logger.js +++ b/utils/logger.js @@ -1,19 +1,30 @@ // module_name: utils_logger_ts // spec_ref: "frontend_state_contract" -const ENABLE_LOG = true; +const ENABLE_DEBUG_LOG = false; +function safe(values) { + return values.map((value) => { + if (value === null || value === undefined || typeof value === 'boolean' || typeof value === 'number') + return value; + if (typeof value === 'string' && /^[a-z0-9_.:/-]{1,96}$/i.test(value)) + return value; + if (value instanceof Error) + return value.name; + return '[redacted]'; + }); +} // @spec_ref frontend_state_contract export function log(scope, ...args) { - if (!ENABLE_LOG) + if (!ENABLE_DEBUG_LOG) return; - console.log(`[ChromeAI:${scope}]`, ...args); + console.log(`[SelectPilot:${scope}]`, ...safe(args)); } // @spec_ref frontend_state_contract export function warn(scope, ...args) { - if (!ENABLE_LOG) + if (!ENABLE_DEBUG_LOG) return; - console.warn(`[ChromeAI:${scope}]`, ...args); + console.warn(`[SelectPilot:${scope}]`, ...safe(args)); } // @spec_ref frontend_state_contract export function error(scope, ...args) { - console.error(`[ChromeAI:${scope}]`, ...args); + console.error(`[SelectPilot:${scope}]`, ...safe(args)); } diff --git a/utils/logger.ts b/utils/logger.ts index 35cd1a2..3330a4c 100644 --- a/utils/logger.ts +++ b/utils/logger.ts @@ -1,20 +1,31 @@ // module_name: utils_logger_ts // spec_ref: "frontend_state_contract" -const ENABLE_LOG = true; +const ENABLE_DEBUG_LOG = false; + +type SafeLogValue = string | number | boolean | null | undefined; + +function safe(values: unknown[]): SafeLogValue[] { + return values.map((value) => { + if (value === null || value === undefined || typeof value === 'boolean' || typeof value === 'number') return value; + if (typeof value === 'string' && /^[a-z0-9_.:/-]{1,96}$/i.test(value)) return value; + if (value instanceof Error) return value.name; + return '[redacted]'; + }); +} // @spec_ref frontend_state_contract export function log(scope: string, ...args: unknown[]): void { - if (!ENABLE_LOG) return; - console.log(`[ChromeAI:${scope}]`, ...args); + if (!ENABLE_DEBUG_LOG) return; + console.log(`[SelectPilot:${scope}]`, ...safe(args)); } // @spec_ref frontend_state_contract export function warn(scope: string, ...args: unknown[]): void { - if (!ENABLE_LOG) return; - console.warn(`[ChromeAI:${scope}]`, ...args); + if (!ENABLE_DEBUG_LOG) return; + console.warn(`[SelectPilot:${scope}]`, ...safe(args)); } // @spec_ref frontend_state_contract export function error(scope: string, ...args: unknown[]): void { - console.error(`[ChromeAI:${scope}]`, ...args); + console.error(`[SelectPilot:${scope}]`, ...safe(args)); } diff --git a/utils/storage.js b/utils/storage.js index f42c5f6..0db16e7 100644 --- a/utils/storage.js +++ b/utils/storage.js @@ -54,6 +54,21 @@ export async function getDecrypted(key) { return null; } } +export async function setEncryptedJSON(key, value) { + await setEncrypted(key, JSON.stringify(value)); +} +export async function getEncryptedJSON(key) { + const decrypted = await getDecrypted(key); + if (decrypted === null) + return null; + try { + return JSON.parse(decrypted); + } + catch { + error('storage', 'decrypted_json_invalid'); + return null; + } +} export async function setJSON(key, value) { await chrome.storage.local.set({ [key]: JSON.stringify(value) }); } diff --git a/utils/storage.ts b/utils/storage.ts index 8706359..59a5f3d 100644 --- a/utils/storage.ts +++ b/utils/storage.ts @@ -56,6 +56,21 @@ export async function getDecrypted(key: string): Promise { } } +export async function setEncryptedJSON(key: string, value: unknown): Promise { + await setEncrypted(key, JSON.stringify(value)); +} + +export async function getEncryptedJSON(key: string): Promise { + const decrypted = await getDecrypted(key); + if (decrypted === null) return null; + try { + return JSON.parse(decrypted) as T; + } catch { + error('storage', 'decrypted_json_invalid'); + return null; + } +} + export async function setJSON(key: string, value: unknown): Promise { await chrome.storage.local.set({ [key]: JSON.stringify(value) }); } From 832d387cd2eb6b4e8f71e9fae6cd258c999c825e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 12:48:32 +0200 Subject: [PATCH 51/66] fix(ci): synchronize hardened content bundle --- content/content-script.bundle.js | 17 ++++++++++++++--- tests/panel/api-request-hardening.test.mjs | 2 ++ tests/panel/local-settings-boundary.test.mjs | 2 ++ 3 files changed, 18 insertions(+), 3 deletions(-) diff --git a/content/content-script.bundle.js b/content/content-script.bundle.js index 6071515..ccccbf8 100644 --- a/content/content-script.bundle.js +++ b/content/content-script.bundle.js @@ -22,11 +22,22 @@ } // utils/logger.js - var ENABLE_LOG = true; + var ENABLE_DEBUG_LOG = false; + function safe(values) { + return values.map((value) => { + if (value === null || value === void 0 || typeof value === "boolean" || typeof value === "number") + return value; + if (typeof value === "string" && /^[a-z0-9_.:/-]{1,96}$/i.test(value)) + return value; + if (value instanceof Error) + return value.name; + return "[redacted]"; + }); + } function log(scope, ...args) { - if (!ENABLE_LOG) + if (!ENABLE_DEBUG_LOG) return; - console.log(`[ChromeAI:${scope}]`, ...args); + console.log(`[SelectPilot:${scope}]`, ...safe(args)); } // content/content-script.ts diff --git a/tests/panel/api-request-hardening.test.mjs b/tests/panel/api-request-hardening.test.mjs index 30b3acc..248ca12 100644 --- a/tests/panel/api-request-hardening.test.mjs +++ b/tests/panel/api-request-hardening.test.mjs @@ -1,3 +1,5 @@ +// module_name: api_request_hardening_tests +// spec_ref: "privacy_and_debug_policy" import assert from 'node:assert/strict'; import test from 'node:test'; diff --git a/tests/panel/local-settings-boundary.test.mjs b/tests/panel/local-settings-boundary.test.mjs index cebd660..3aefbd8 100644 --- a/tests/panel/local-settings-boundary.test.mjs +++ b/tests/panel/local-settings-boundary.test.mjs @@ -1,3 +1,5 @@ +// module_name: local_settings_boundary_tests +// spec_ref: "privacy_and_debug_policy" import assert from 'node:assert/strict'; import { readFile } from 'node:fs/promises'; import test from 'node:test'; From d6174525ea1ea8c7bb96ca4acf63ca64aec74dd5 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 12:48:44 +0200 Subject: [PATCH 52/66] fix(billing): bound hostile checkout and authority inputs Intent: AXIOM-20260830-OXBH25 Reject malformed and oversized authority requests, enforce JSON media types, bound server sockets, remove checkout HTML injection, and cap polling/network waits. Not run: full verification executes on Hetzner. --- services/entitlement-authority/server.mjs | 52 +++++++++++---- site/checkout.js | 75 +++++++++++++++++----- tests/panel/checkout-hardening.test.mjs | 19 ++++++ tests/panel/entitlement-authority.test.mjs | 18 ++++++ 4 files changed, 136 insertions(+), 28 deletions(-) create mode 100644 tests/panel/checkout-hardening.test.mjs diff --git a/services/entitlement-authority/server.mjs b/services/entitlement-authority/server.mjs index c546006..5f51085 100644 --- a/services/entitlement-authority/server.mjs +++ b/services/entitlement-authority/server.mjs @@ -9,6 +9,25 @@ import { mkdir } from 'node:fs/promises'; const TIERS = new Set(['essential', 'plus', 'pro']); const MAX_BODY_BYTES = 256 * 1024; const PADDLE_TOLERANCE_SECONDS = 5; +const MAX_IDENTIFIER_CHARS = 200; + +function parseObject(rawBody) { + let parsed; + try { + parsed = JSON.parse(rawBody || '{}'); + } catch { + throw Object.assign(new Error('invalid_json'), { status: 400 }); + } + if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) { + throw Object.assign(new Error('invalid_json_object'), { status: 400 }); + } + return parsed; +} + +function identifier(value) { + return typeof value === 'string' && value.trim().length > 0 && value.trim().length <= MAX_IDENTIFIER_CHARS + ? value.trim() : null; +} function required(name, env = process.env) { const value = env[name]?.trim(); @@ -143,7 +162,7 @@ export async function createAuthority(config = {}) { return current; }; - return createServer(async (req, res) => { + const server = createServer(async (req, res) => { try { if (req.headers.origin === allowedWebOrigin) { res.setHeader('access-control-allow-origin', allowedWebOrigin); @@ -157,10 +176,13 @@ export async function createAuthority(config = {}) { } if (req.method === 'GET' && req.url === '/health') return send(res, 200, { status: 'ok' }); if (req.method !== 'POST') return send(res, 404, { error: 'not_found' }); + if (!String(req.headers['content-type'] || '').toLowerCase().startsWith('application/json')) { + return send(res, 415, { error: 'json_content_type_required' }); + } const rawBody = await readBody(req); if (req.url === '/v1/entitlements/verify') { - const token = JSON.parse(rawBody || '{}').token?.trim(); + const token = identifier(parseObject(rawBody).token); if (!token) return send(res, 400, { error: 'token_required' }); const state = await loadState(statePath); const record = state.tokens[tokenHash(token)]; @@ -171,8 +193,8 @@ export async function createAuthority(config = {}) { } if (req.url === '/v1/trials/start') { - const installationId = JSON.parse(rawBody || '{}').installation_id?.trim(); - if (!installationId || installationId.length > 200) return send(res, 400, { error: 'installation_id_required' }); + const installationId = identifier(parseObject(rawBody).installation_id); + if (!installationId) return send(res, 400, { error: 'installation_id_required' }); let response; await serial(async () => { const state = await loadState(statePath); @@ -196,7 +218,7 @@ export async function createAuthority(config = {}) { } if (req.url === '/v1/claims/redeem') { - const claimId = JSON.parse(rawBody || '{}').claim_id?.trim(); + const claimId = identifier(parseObject(rawBody).claim_id); if (!claimId) return send(res, 400, { error: 'claim_id_required' }); let response = { status: 'pending' }; await serial(async () => { @@ -219,7 +241,7 @@ export async function createAuthority(config = {}) { } if (req.url === '/v1/claims/ack') { - const claimId = JSON.parse(rawBody || '{}').claim_id?.trim(); + const claimId = identifier(parseObject(rawBody).claim_id); if (!claimId) return send(res, 400, { error: 'claim_id_required' }); await serial(async () => { const state = await loadState(statePath); @@ -236,14 +258,15 @@ export async function createAuthority(config = {}) { if (!verifyPaddleSignature(rawBody, req.headers['paddle-signature'], webhookSecret)) { return send(res, 401, { error: 'invalid_webhook_signature' }); } - const event = JSON.parse(rawBody); - const claimId = event?.data?.custom_data?.claim_id; + const event = parseObject(rawBody); + const claimId = identifier(event?.data?.custom_data?.claim_id); const product = purchasedItem(event, priceMap); - const subjectId = event?.data?.subscription_id || event?.data?.id; - if (!event?.event_id || !subjectId) return send(res, 422, { error: 'unsupported_event' }); + const subjectId = identifier(event?.data?.subscription_id || event?.data?.id); + const eventId = identifier(event?.event_id); + if (!eventId || !subjectId) return send(res, 422, { error: 'unsupported_event' }); await serial(async () => { const state = await loadState(statePath); - if (state.events[event.event_id]) return; + if (state.events[eventId]) return; if (['transaction.completed', 'subscription.activated'].includes(event.event_type)) { if (!claimId || !product) throw Object.assign(new Error('unsupported_event'), { status: 422 }); const previousHash = state.subjects[subjectId]; @@ -271,7 +294,7 @@ export async function createAuthority(config = {}) { const currentHash = state.subjects[subjectId]; if (currentHash && state.tokens[currentHash]) state.tokens[currentHash].revokedAt = Date.now(); } - state.events[event.event_id] = { handledAt: Date.now(), type: event.event_type }; + state.events[eventId] = { handledAt: Date.now(), type: event.event_type }; await saveState(statePath, state); }); return send(res, 200, { accepted: true }); @@ -290,6 +313,11 @@ export async function createAuthority(config = {}) { send(res, status, { error: error.status ? error.message : 'internal_error' }); } }); + server.requestTimeout = 15_000; + server.headersTimeout = 10_000; + server.keepAliveTimeout = 5_000; + server.maxRequestsPerSocket = 100; + return server; } if (process.argv[1] && resolve(process.argv[1]) === resolve(new URL(import.meta.url).pathname)) { diff --git a/site/checkout.js b/site/checkout.js index 7e81a95..1f94bc5 100644 --- a/site/checkout.js +++ b/site/checkout.js @@ -8,31 +8,74 @@ const claimStorageKey = `selectpilot_claim_${tier || 'unknown'}`; const claimId = localStorage.getItem(claimStorageKey) || crypto.randomUUID(); localStorage.setItem(claimStorageKey, claimId); + const POLL_DEADLINE_MS = 5 * 60 * 1000; + const REQUEST_TIMEOUT_MS = 10_000; + const startedAt = Date.now(); + let pollAttempt = 0; - function message(title, detail, action = '') { - state.innerHTML = `

    ${title}

    ${detail}

    ${action}
    `; + function message(title, detail, action) { + const container = document.createElement('div'); + const heading = document.createElement('h2'); + const paragraph = document.createElement('p'); + heading.textContent = title; + paragraph.textContent = detail; + container.append(heading, paragraph); + if (action) container.append(action); + state.replaceChildren(container); + } + + async function post(path, body) { + const controller = new AbortController(); + const timeout = setTimeout(() => controller.abort(), REQUEST_TIMEOUT_MS); + try { + const response = await fetch(`${config.entitlementAuthorityUrl}${path}`, { + method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify(body), + cache: 'no-store', signal: controller.signal, + }); + if (!response.ok) throw new Error(`authority_http_${response.status}`); + return await response.json(); + } finally { + clearTimeout(timeout); + } + } + + function scheduleRedeem() { + if (Date.now() - startedAt >= POLL_DEADLINE_MS) { + message('License preparation needs attention', 'Reload this page to check again. Your payment has not been repeated.'); + return; + } + const delay = Math.min(15_000, 1_500 * (2 ** Math.min(pollAttempt, 4))); + pollAttempt += 1; + setTimeout(() => redeem().catch(scheduleRedeem), delay); } async function redeem() { - const response = await fetch(`${config.entitlementAuthorityUrl}/v1/claims/redeem`, { - method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ claim_id: claimId }) - }); - const result = await response.json(); + const result = await post('/v1/claims/redeem', { claim_id: claimId }); if (result.status === 'ready' && result.token) { const token = result.token; - message('Your license is ready', 'Copy this key, then enter it in SelectPilot.', `

    ${token}

    `); - document.querySelector('#copy-license').addEventListener('click', async () => { - await navigator.clipboard.writeText(token); - await fetch(`${config.entitlementAuthorityUrl}/v1/claims/ack`, { - method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ claim_id: claimId }) - }); - localStorage.removeItem(claimStorageKey); - message('License copied', 'Return to SelectPilot and choose Activate.'); + const action = document.createElement('div'); + const key = document.createElement('code'); + const button = document.createElement('button'); + key.textContent = token; + button.className = 'button'; + button.type = 'button'; + button.textContent = 'Copy license key'; + action.append(key, button); + message('Your license is ready', 'Copy this key, then enter it in SelectPilot.', action); + button.addEventListener('click', async () => { + try { + await navigator.clipboard.writeText(token); + await post('/v1/claims/ack', { claim_id: claimId }); + localStorage.removeItem(claimStorageKey); + message('License copied', 'Return to SelectPilot and choose Activate.'); + } catch { + message('License copied, acknowledgement pending', 'Keep this page open and try Copy license key again before closing it.', action); + } }); return; } message('Payment received', 'Your license is being prepared. This page will update automatically.'); - setTimeout(redeem, 1800); + scheduleRedeem(); } if (!config || !priceId || !window.Paddle || !/^pri_/.test(priceId)) { @@ -41,7 +84,7 @@ } if (config.paddleEnvironment === 'sandbox') Paddle.Environment.set('sandbox'); Paddle.Initialize({ token: config.paddleClientToken, eventCallback(event) { - if (event.name === 'checkout.completed') redeem().catch(() => message('Payment received', 'Keep this page open while your license is prepared.')); + if (event.name === 'checkout.completed') redeem().catch(scheduleRedeem); }}); Paddle.Checkout.open({ items: [{ priceId, quantity: 1 }], customData: { claim_id: claimId }, settings: { displayMode: 'inline', frameTarget: 'checkout-state', frameInitialHeight: 450, frameStyle: 'width:100%;border:0;background:transparent;' } }); })(); diff --git a/tests/panel/checkout-hardening.test.mjs b/tests/panel/checkout-hardening.test.mjs new file mode 100644 index 0000000..9131c52 --- /dev/null +++ b/tests/panel/checkout-hardening.test.mjs @@ -0,0 +1,19 @@ +import assert from 'node:assert/strict'; +import { readFile } from 'node:fs/promises'; +import test from 'node:test'; + +test('checkout renders untrusted values as text and bounds polling/network waits', async () => { + const source = await readFile(new URL('../../site/checkout.js', import.meta.url), 'utf8'); + assert.doesNotMatch(source, /innerHTML/); + assert.match(source, /textContent = token/); + assert.match(source, /AbortController/); + assert.match(source, /POLL_DEADLINE_MS/); + assert.match(source, /Math\.min\(15_000/); +}); + +test('checkout never reports final success before claim acknowledgement', async () => { + const source = await readFile(new URL('../../site/checkout.js', import.meta.url), 'utf8'); + const acknowledge = source.indexOf("await post('/v1/claims/ack'"); + const success = source.indexOf("message('License copied'"); + assert.ok(acknowledge >= 0 && success > acknowledge); +}); diff --git a/tests/panel/entitlement-authority.test.mjs b/tests/panel/entitlement-authority.test.mjs index 9dc6ea9..f72f551 100644 --- a/tests/panel/entitlement-authority.test.mjs +++ b/tests/panel/entitlement-authority.test.mjs @@ -132,3 +132,21 @@ test('Paddle signature supports rotation and rejects stale or forged events', () assert.equal(verifyPaddleSignature(body, `ts=${now};ts=${now};h1=${valid}`, 'secret', now), false); assert.equal(verifyPaddleSignature(body, `ts=${now};h1=malformed;h1=${valid}`, 'secret', now), false); }); + +test('authority rejects malformed, mistyped, oversized and ambiguous identifiers explicitly', async (t) => { + const base = await fixture(t); + assert.equal((await post(base, '/v1/trials/start', '{')).status, 400); + assert.equal((await fetch(`${base}/v1/trials/start`, { method: 'POST', body: '{}' })).status, 415); + assert.equal((await post(base, '/v1/trials/start', [])).status, 400); + assert.equal((await post(base, '/v1/trials/start', { installation_id: 'x'.repeat(201) })).status, 400); + assert.equal((await post(base, '/v1/claims/redeem', { claim_id: 'x'.repeat(201) })).status, 400); + assert.equal((await post(base, '/v1/entitlements/verify', { token: 'x'.repeat(201) })).status, 400); +}); + +test('concurrent trial abuse preserves one deterministic entitlement', async (t) => { + const base = await fixture(t); + const responses = await Promise.all(Array.from({ length: 200 }, () => + post(base, '/v1/trials/start', { installation_id: 'same-installation' }).then((response) => response.json()) + )); + assert.equal(new Set(responses.map((entry) => entry.token)).size, 1); +}); From 262497ada51a4c1da8fe15d038f16e2cd0252f49 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 12:50:22 +0200 Subject: [PATCH 53/66] fix(runtime): bound logs streams and error disclosure Intent: AXIOM-20260830-OXBH25 Normalize hostile trace identifiers, recursively redact runtime metadata, rotate feedback logs, cap SSE streams, and suppress raw model errors. Not run: full verification executes on Hetzner. --- server/nano_server.py | 25 +++++++++++++++--------- tests/server/test_validation_pipeline.py | 6 +++++- 2 files changed, 21 insertions(+), 10 deletions(-) diff --git a/server/nano_server.py b/server/nano_server.py index 5d5a954..fbce321 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -406,10 +406,9 @@ def enforce_contract_fields(payload: dict, contract: OperationContract) -> None: def _resolve_trace_id(payload: dict, headers: dict[str, str]) -> str: header_trace = str(headers.get("x-selectpilot-trace-id") or headers.get("x-trace-id") or "").strip() body_trace = str(payload.get("session_id") or "").strip() if isinstance(payload, dict) else "" - if header_trace: - return header_trace - if body_trace: - return body_trace + candidate = header_trace or body_trace + if candidate and re.fullmatch(r"[A-Za-z0-9_.:-]{1,64}", candidate): + return candidate basis = { "payload": payload if isinstance(payload, dict) else {}, "content_type": str(headers.get("content-type") or ""), @@ -461,6 +460,8 @@ def write_port_info(port_file: Path, port: int): RUNTIME_REGISTRY_PATH = RUNTIME_DIR / "model_registry.runtime.json" LIVE_FEEDBACK_PATH = RUNTIME_DIR / "live_feedback.jsonl" _LIVE_FEEDBACK_LOCK = threading.Lock() +MAX_LIVE_FEEDBACK_BYTES = 2 * 1024 * 1024 +MAX_RUNTIME_META_STREAMS = 4 RUNTIME_META_EVENT_VERSION = "1.0" RUNTIME_META_DEFAULT_LATENCY_HINT_MS = 1200 @@ -496,6 +497,9 @@ def _append_live_feedback(event: dict) -> None: } line = json.dumps(entry, ensure_ascii=False, separators=(",", ":")) with _LIVE_FEEDBACK_LOCK: + if LIVE_FEEDBACK_PATH.exists() and LIVE_FEEDBACK_PATH.stat().st_size >= MAX_LIVE_FEEDBACK_BYTES: + rotated = LIVE_FEEDBACK_PATH.with_suffix(".jsonl.previous") + os.replace(LIVE_FEEDBACK_PATH, rotated) with LIVE_FEEDBACK_PATH.open("a", encoding="utf-8") as handle: handle.write(f"{line}\n") except Exception: @@ -753,7 +757,7 @@ def sanitize_runtime_meta_details(details: dict | None) -> dict: elif isinstance(value, list): safe[key_str] = len(value) elif isinstance(value, dict): - safe[key_str] = {k: v for k, v in value.items() if isinstance(v, (str, int, float, bool, type(None)))} + safe[key_str] = sanitize_runtime_meta_details(value) else: safe[key_str] = str(value) return safe @@ -1494,6 +1498,9 @@ def do_GET(self): }) return if path == '/runtime-meta/stream': + if RUNTIME_META.active_stream_count() >= MAX_RUNTIME_META_STREAMS: + self._write_error(429, "runtime_meta_stream_limit", "Too many active runtime streams") + return query = parse_qs(parsed.query) after = query.get("after", [""])[0] header_last_event = str(self.headers.get("Last-Event-ID") or "").strip() @@ -1831,7 +1838,7 @@ def _execute_summarize(): trace_id, contract.name, "runtime_error", - {"code": "ollama_unavailable", "message": str(e)}, + {"code": "ollama_unavailable"}, ) emit_runtime_meta( event_type="STEP_FAILED", @@ -1839,7 +1846,7 @@ def _execute_summarize(): operation=contract.name, status="error", step=current_step, - message=str(e), + message="Local model runtime is unavailable", details={"code": "ollama_unavailable"}, ) emit_runtime_meta( @@ -1847,11 +1854,11 @@ def _execute_summarize(): trace_id=trace_id, operation=contract.name, status="error", - message=str(e), + message="Local model runtime is unavailable", duration_ms=duration_ms, details={"code": "ollama_unavailable"}, ) - self._write_error(503, "ollama_unavailable", str(e)) + self._write_error(503, "ollama_unavailable", "Local model runtime is unavailable") return if isinstance(resp, dict): diff --git a/tests/server/test_validation_pipeline.py b/tests/server/test_validation_pipeline.py index c977185..f7944fb 100644 --- a/tests/server/test_validation_pipeline.py +++ b/tests/server/test_validation_pipeline.py @@ -211,6 +211,9 @@ def test_trace_id_resolves_from_header_then_body_then_generated(self) -> None: generated = _resolve_trace_id({}, {}) self.assertTrue(isinstance(generated, str) and len(generated) > 0) + hostile = _resolve_trace_id({}, {"x-selectpilot-trace-id": "secret\n" + "x" * 200}) + self.assertRegex(hostile, r"^sp_[0-9a-f]{24}$") + def test_runtime_meta_details_are_sanitized_for_privacy(self) -> None: sanitized = sanitize_runtime_meta_details( { @@ -218,13 +221,14 @@ def test_runtime_meta_details_are_sanitized_for_privacy(self) -> None: "text": "secret", "prompt": "never log me", "safe_scalar": "ok", - "nested": {"allowed": "yes", "blocked": ["x"]}, + "nested": {"allowed": "yes", "prompt": "nested secret", "blocked": ["x"]}, } ) self.assertNotIn("text", sanitized) self.assertNotIn("prompt", sanitized) self.assertEqual(sanitized.get("safe_scalar"), "ok") self.assertEqual(sanitized.get("request_fields"), 2) + self.assertNotIn("prompt", sanitized.get("nested", {})) def test_runtime_meta_event_contract_contains_privacy_flags(self) -> None: event = build_runtime_meta_event( From 0078352994cdfb4b95720618df497c55a5cf4644 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 12:52:22 +0200 Subject: [PATCH 54/66] fix(ci): map checkout hardening coverage --- tests/panel/checkout-hardening.test.mjs | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tests/panel/checkout-hardening.test.mjs b/tests/panel/checkout-hardening.test.mjs index 9131c52..85d1e7e 100644 --- a/tests/panel/checkout-hardening.test.mjs +++ b/tests/panel/checkout-hardening.test.mjs @@ -1,3 +1,5 @@ +// module_name: checkout_hardening_tests +// spec_ref: "privacy_and_debug_policy" import assert from 'node:assert/strict'; import { readFile } from 'node:fs/promises'; import test from 'node:test'; From 1e7b53ba3270aae6dc6b0d2224718977ecdf9024 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 13:01:34 +0200 Subject: [PATCH 55/66] fix(governance): verify signed Axiom intent Intent: AXIOM-20260830-OXBH25 --- .github/workflows/axiom.yml | 30 +++++++-- governance/axiom-trust.json | 7 ++ package.json | 1 + scripts/verify-axiom-intent.mjs | 104 ++++++++++++++++++++++++++++++ tests/panel/axiom-intent.test.mjs | 28 ++++++++ 5 files changed, 165 insertions(+), 5 deletions(-) create mode 100644 governance/axiom-trust.json create mode 100644 scripts/verify-axiom-intent.mjs create mode 100644 tests/panel/axiom-intent.test.mjs diff --git a/.github/workflows/axiom.yml b/.github/workflows/axiom.yml index cfc26c4..c57a57d 100644 --- a/.github/workflows/axiom.yml +++ b/.github/workflows/axiom.yml @@ -11,8 +11,28 @@ permissions: pull-requests: read jobs: - axiom: - uses: kristoffersodersten/axiom-core/.github/workflows/axiom-downstream.yml@main - with: - intent_path: intent.json - strict: true + axiom_validate: + runs-on: ubuntu-latest + timeout-minutes: 5 + steps: + - name: Checkout downstream repository + uses: actions/checkout@v6 + with: + fetch-depth: 0 + + - name: Setup Node + uses: actions/setup-node@v6 + with: + node-version: 24 + + - name: Load changed files + run: | + mkdir -p .axiom-runtime + if [ "${{ github.event_name }}" = "pull_request" ]; then + git diff --name-only "origin/${{ github.base_ref }}"...HEAD > .axiom-runtime/changed-files.txt + else + git diff --name-only "${{ github.event.before }}" "${{ github.sha }}" > .axiom-runtime/changed-files.txt + fi + + - name: Verify signed Axiom intent and change boundary + run: node scripts/verify-axiom-intent.mjs --changed-files .axiom-runtime/changed-files.txt diff --git a/governance/axiom-trust.json b/governance/axiom-trust.json new file mode 100644 index 0000000..94600e8 --- /dev/null +++ b/governance/axiom-trust.json @@ -0,0 +1,7 @@ +{ + "schema_version": 1, + "authority": "Namaka-Inc/axiom-core", + "authority_sha": "cd12d0547d210ea7e9d985fce55f90acc4d6317a", + "policy_sha256": "932bbb231cb5e1aeb35d6bbe6327db6f3b1b3cf1e125e38366cbccbbc59434cf", + "owner_public_key_sha256": "179b5520ccd5f32f04f0fc09df59a619975a2f1c93b24674883b07fa0ce301b8" +} diff --git a/package.json b/package.json index 5b30f72..6fadbe2 100644 --- a/package.json +++ b/package.json @@ -24,6 +24,7 @@ "runtime:rollback": "node ./scripts/runtime-rollback.mjs", "runtime:show-policy": "node ./scripts/runtime-show-policy.mjs", "verify:admission": "node ./scripts/verify-code-admission.mjs", + "verify:axiom-intent": "node ./scripts/verify-axiom-intent.mjs", "sync:manifest-version": "node ./scripts/sync-manifest-version.mjs", "build": "pnpm sync:manifest-version && node ./scripts/generate-extraction-presets.mjs && tsc -p tsconfig.json && node ./scripts/build-extension.mjs", "lint": "node ./scripts/generate-extraction-presets.mjs --check && eslint --max-warnings=0 .", diff --git a/scripts/verify-axiom-intent.mjs b/scripts/verify-axiom-intent.mjs new file mode 100644 index 0000000..7b87829 --- /dev/null +++ b/scripts/verify-axiom-intent.mjs @@ -0,0 +1,104 @@ +// module_name: axiom_intent_verifier +// spec_ref: "validation_layer" +import { createHash, createPublicKey, verify } from 'node:crypto'; +import { readFile } from 'node:fs/promises'; +import { resolve } from 'node:path'; + +const SIGNED_FIELDS = [ + 'id', + 'goal', + 'linear_issue_id', + 'canonical_reference', + 'contract_target', + 'layer', + 'acceptance_criteria', + 'verification_method', + 'constraints', + 'allowed_actions', + 'forbidden_actions', + 'perception_targets', + 'expires_at', + 'nonce', +]; + +function stableStringify(value) { + if (value === null || typeof value !== 'object') return JSON.stringify(value); + if (Array.isArray(value)) return `[${value.map(stableStringify).join(',')}]`; + return `{${Object.keys(value).sort().map((key) => `${JSON.stringify(key)}:${stableStringify(value[key])}`).join(',')}}`; +} + +function matchesPattern(file, pattern) { + let expression = '^'; + for (let index = 0; index < pattern.length; index += 1) { + const character = pattern[index]; + if (character === '*' && pattern[index + 1] === '*') { + if (pattern[index + 2] === '/') { + expression += '(?:.*/)?'; + index += 2; + } else { + expression += '.*'; + index += 1; + } + } else if (character === '*') { + expression += '[^/]*'; + } else { + expression += character.replace(/[|\\{}()[\]^$+?.]/g, '\\$&'); + } + } + return new RegExp(`${expression}$`).test(file); +} + +// @spec_ref validation_layer +export function verifyAxiomIntent(intent, trust, { changedFiles = [], nowSeconds = Math.floor(Date.now() / 1000) } = {}) { + const failures = []; + const publicKeyBytes = Buffer.from(intent?.signature?.public_key || '', 'base64'); + const publicKeyHash = createHash('sha256').update(publicKeyBytes).digest('hex'); + + if (intent?.signature?.algorithm !== 'ed25519') failures.push('invalid_algorithm'); + if (intent?.signature?.public_key !== intent?.signer?.public_key) failures.push('signer_key_mismatch'); + if (intent?.signer?.role !== 'owner') failures.push('invalid_signer_role'); + if (publicKeyHash !== trust?.owner_public_key_sha256) failures.push('untrusted_signer'); + if (JSON.stringify(intent?.signature?.signed_payload_fields) !== JSON.stringify(SIGNED_FIELDS)) failures.push('signed_fields_mismatch'); + if (!Number.isInteger(intent?.expires_at) || intent.expires_at <= nowSeconds) failures.push('intent_expired'); + if (!/^AXIOM-[A-Z0-9-]+$/.test(intent?.id || '')) failures.push('invalid_intent_id'); + if (!/^SOD-[0-9]+$/.test(intent?.linear_issue_id || '')) failures.push('invalid_linear_issue'); + if (!/^kristoffersodersten\/selectpilot@[0-9a-f]{40}$/.test(intent?.canonical_reference || '')) failures.push('invalid_canonical_reference'); + + try { + const payload = Object.fromEntries(SIGNED_FIELDS.map((field) => [field, intent[field]])); + const publicKey = createPublicKey({ key: publicKeyBytes, format: 'der', type: 'spki' }); + const valid = verify( + null, + Buffer.from(stableStringify(payload), 'utf8'), + publicKey, + Buffer.from(intent.signature.value || '', 'base64'), + ); + if (!valid) failures.push('invalid_signature'); + } catch { + failures.push('invalid_signature'); + } + + for (const file of changedFiles) { + if (intent.forbidden_actions?.some((pattern) => matchesPattern(file, pattern))) { + failures.push(`forbidden_file_change:${file}`); + } else if (!intent.allowed_actions?.some((pattern) => matchesPattern(file, pattern))) { + failures.push(`unknown_file_change:${file}`); + } + } + + return { ok: failures.length === 0, failures }; +} + +async function main() { + const changedFilesIndex = process.argv.indexOf('--changed-files'); + const changedFiles = changedFilesIndex >= 0 + ? (await readFile(resolve(process.argv[changedFilesIndex + 1]), 'utf8')).split(/\r?\n/).filter(Boolean) + : []; + const intent = JSON.parse(await readFile(resolve('intent.json'), 'utf8')); + const trust = JSON.parse(await readFile(resolve('governance/axiom-trust.json'), 'utf8')); + const result = verifyAxiomIntent(intent, trust, { changedFiles }); + process.stdout.write(`${JSON.stringify({ ...result, intent_id: intent.id, authority_sha: trust.authority_sha })}\n`); + if (!result.ok) process.exitCode = 1; +} + +if (process.argv[1] && resolve(process.argv[1]) === resolve(new URL(import.meta.url).pathname)) await main(); diff --git a/tests/panel/axiom-intent.test.mjs b/tests/panel/axiom-intent.test.mjs new file mode 100644 index 0000000..9d769b1 --- /dev/null +++ b/tests/panel/axiom-intent.test.mjs @@ -0,0 +1,28 @@ +// module_name: axiom_intent_tests +// spec_ref: "testing_strategy.integration_tests" +import assert from 'node:assert/strict'; +import { readFile } from 'node:fs/promises'; +import test from 'node:test'; +import { verifyAxiomIntent } from '../../scripts/verify-axiom-intent.mjs'; + +const intent = JSON.parse(await readFile(new URL('../../intent.json', import.meta.url), 'utf8')); +const trust = JSON.parse(await readFile(new URL('../../governance/axiom-trust.json', import.meta.url), 'utf8')); + +test('Axiom intent is owner-signed, trusted, current, and bounds changed files', () => { + const result = verifyAxiomIntent(intent, trust, { + changedFiles: ['background/background.ts', 'tests/panel/axiom-intent.test.mjs'], + nowSeconds: 1_788_000_000, + }); + assert.deepEqual(result, { ok: true, failures: [] }); +}); + +test('Axiom intent verification fails closed on tampering, expiry, and forbidden files', () => { + const tampered = structuredClone(intent); + tampered.goal = 'different goal'; + assert.ok(verifyAxiomIntent(tampered, trust, { nowSeconds: 1_788_000_000 }).failures.includes('invalid_signature')); + assert.ok(verifyAxiomIntent(intent, trust, { nowSeconds: intent.expires_at }).failures.includes('intent_expired')); + assert.ok(verifyAxiomIntent(intent, trust, { + changedFiles: ['secrets/private.key'], + nowSeconds: 1_788_000_000, + }).failures.includes('forbidden_file_change:secrets/private.key')); +}); From 46bd07b4e51c82e2c0df8184855310d297dc69f6 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Sun, 30 Aug 2026 13:07:52 +0200 Subject: [PATCH 56/66] fix(macos): preserve hardware task routing Intent: AXIOM-20260830-OXBH25 --- .../com.selectpilot.bridge.plist.template | 4 +++ installer/macos/scripts/postinstall | 16 ++++++++-- tests/panel/bootstrap-contract.test.mjs | 30 +++++++++++++++++++ 3 files changed, 48 insertions(+), 2 deletions(-) mode change 100644 => 100755 installer/macos/scripts/postinstall diff --git a/installer/macos/com.selectpilot.bridge.plist.template b/installer/macos/com.selectpilot.bridge.plist.template index 582c9d8..e547e77 100644 --- a/installer/macos/com.selectpilot.bridge.plist.template +++ b/installer/macos/com.selectpilot.bridge.plist.template @@ -19,9 +19,13 @@ CHROMEAI_EXPECTED_PORT_FILE__RUN_DIR__/port.info CHROMEAI_ALLOWED_RANGE8083 CHROMEAI_OLLAMA_BASE_URL__OLLAMA_BASE_URL__ + CHROMEAI_RUNTIME_PROFILE__RUNTIME_PROFILE__ CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ + CHROMEAI_OLLAMA_FAST_MODEL__OLLAMA_FAST_MODEL__ CHROMEAI_OLLAMA_EMBED_MODEL__OLLAMA_EMBED_MODEL__ CHROMEAI_OLLAMA_NUM_CTX__OLLAMA_NUM_CTX__ + CHROMEAI_OLLAMA_FAST_NUM_CTX__OLLAMA_FAST_NUM_CTX__ + CHROMEAI_MAX_INPUT_CHARS__MAX_INPUT_CHARS__ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_AUTO_PULL_MODELS0 diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall old mode 100644 new mode 100755 index a12990e..99ba646 --- a/installer/macos/scripts/postinstall +++ b/installer/macos/scripts/postinstall @@ -32,10 +32,18 @@ import sys sys.path.insert(0, sys.argv[1]) from runtime_profiles import get_runtime_profile, recommend_runtime_profile profile = get_runtime_profile(recommend_runtime_profile()["recommended_profile"]) -print(f"{profile.generation_model}|{profile.embedding_model}|{profile.num_ctx}") +print("|".join(str(value) for value in ( + profile.key, + profile.generation_model, + profile.fast_generation_model, + profile.embedding_model, + profile.num_ctx, + profile.fast_num_ctx, + profile.max_input_chars, +))) PY )" -IFS='|' read -r OLLAMA_MODEL OLLAMA_EMBED_MODEL OLLAMA_NUM_CTX < "$AGENT_PATH" diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index e20b3d4..80e2e33 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -85,3 +85,33 @@ test('production macOS package also binds launchd to Python 3.9 or newer', () => assert.match(launchAgent, /__PYTHON_BIN__/); assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); }); + +test('production macOS package preserves hardware-aware task routing', () => { + const postinstall = readFileSync('./installer/macos/scripts/postinstall', 'utf8'); + const launchAgent = readFileSync('./installer/macos/com.selectpilot.bridge.plist.template', 'utf8'); + + for (const field of [ + 'profile.key', + 'profile.generation_model', + 'profile.fast_generation_model', + 'profile.embedding_model', + 'profile.num_ctx', + 'profile.fast_num_ctx', + 'profile.max_input_chars', + ]) { + assert.ok(postinstall.includes(field), `missing profile field ${field}`); + } + + for (const placeholder of [ + '__RUNTIME_PROFILE__', + '__OLLAMA_MODEL__', + '__OLLAMA_FAST_MODEL__', + '__OLLAMA_EMBED_MODEL__', + '__OLLAMA_NUM_CTX__', + '__OLLAMA_FAST_NUM_CTX__', + '__MAX_INPUT_CHARS__', + ]) { + assert.ok(postinstall.includes(placeholder), `postinstall does not bind ${placeholder}`); + assert.ok(launchAgent.includes(placeholder), `LaunchAgent does not expose ${placeholder}`); + } +}); From ae9ef22893eefd55f02b78be3d4ed2f89d9cb6e0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 00:05:54 +0200 Subject: [PATCH 57/66] fix(release): enforce exact production runtime truth Fail closed on missing Store identity, runtime integrity failures, mutable state errors, and local inference saturation. Preserve exact model routing through the extension UI. Verified: pnpm lint; pnpm test; pnpm test:e2e; pnpm build; pnpm test:privacy; pnpm test:billing; pnpm bench:frontier; runtime policy validation; stress runner. --- .github/workflows/chrome-store-release.yml | 45 +++++ .github/workflows/ci.yml | 23 +-- .gitignore | 1 + README.md | 2 +- ZERO_LEAKAGE.md | 2 +- agent/agent-client.js | 6 +- agent/agent-client.ts | 6 +- agent/agent-pipeline.js | 6 +- agent/agent-pipeline.ts | 6 +- agent/agent-types.ts | 8 + api/nano-client.js | 12 +- api/nano-client.ts | 31 ++- docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md | 4 +- docs/ENTITLEMENT_DEPLOYMENT.md | 4 +- .../com.selectpilot.bridge.plist.template | 3 + installer/macos/scripts/postinstall | 4 +- launchd/com.chromeai.nano.plist | 3 + package.json | 2 +- panel/panel.css | 2 +- panel/panel.html | 8 + panel/panel.js | 89 +++++---- panel/panel.ts | 93 +++++---- pricing/entitlement-public-keys.json | 9 +- scripts/build-extension.mjs | 26 --- scripts/entitlement-keyring.mjs | 78 ++++++++ scripts/install-entitlement-systemd.sh | 2 + scripts/install-macos-local.sh | 9 +- scripts/package-chrome-store.mjs | 65 +++++-- scripts/preflight-entitlement-deployment.sh | 3 + scripts/spec-compiler.mjs | 2 +- scripts/verify-entitlement-key.mjs | 6 +- server/nano_server.py | 182 +++++++++++++++--- services/entitlement-authority/.env.example | 1 + tests/e2e/extension-first-run-flow.spec.mjs | 10 + tests/e2e/panel-harness.html | 14 +- tests/e2e/playwright.config.mjs | 2 +- tests/e2e/privacy-and-runtime.spec.mjs | 3 +- tests/panel/bootstrap-contract.test.mjs | 4 + tests/panel/capability-truth.test.mjs | 2 + tests/panel/store-package.test.mjs | 48 ++++- tests/server/test_capability_truth.py | 50 +++++ 41 files changed, 674 insertions(+), 202 deletions(-) create mode 100644 .github/workflows/chrome-store-release.yml create mode 100644 scripts/entitlement-keyring.mjs diff --git a/.github/workflows/chrome-store-release.yml b/.github/workflows/chrome-store-release.yml new file mode 100644 index 0000000..5172f41 --- /dev/null +++ b/.github/workflows/chrome-store-release.yml @@ -0,0 +1,45 @@ +name: Chrome Web Store release candidate + +on: + workflow_dispatch: + +permissions: + contents: read + +jobs: + package: + environment: chrome-web-store-production + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v7 + - uses: pnpm/action-setup@v6 + with: + version: 10.33.0 + - uses: actions/setup-node@v7 + with: + node-version: 22 + cache: pnpm + cache-dependency-path: pnpm-lock.yaml + - uses: actions/setup-python@v7 + with: + python-version: "3.12" + - name: Install dependencies + run: pnpm install --frozen-lockfile + - name: Install local billing runtime dependencies + run: python -m pip install --requirement requirements-billing.txt + - name: Verify source and runtime contracts + run: pnpm lint && pnpm typecheck && pnpm test && pnpm test:privacy + - name: Package authorized Store candidate + env: + SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON: ${{ secrets.SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON }} + SELECTPILOT_SOURCE_SHA: ${{ github.sha }} + run: pnpm package:store + - name: Upload exact-revision Store candidate + uses: actions/upload-artifact@v7 + with: + name: selectpilot-chrome-web-store-${{ github.sha }} + path: | + dist/chrome-web-store/*.zip + dist/chrome-web-store/*.zip.json + if-no-files-found: error + retention-days: 30 diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index fd179e1..4179aee 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -95,8 +95,16 @@ jobs: - name: Validate Chrome Web Store assets run: pnpm validate:store - - name: Package Chrome Web Store release candidate - run: pnpm package:store + - name: Verify Store packaging fails closed without production identity + env: + SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON: "" + run: | + if node scripts/package-chrome-store.mjs >store-package.log 2>&1; then + echo "Store packaging unexpectedly succeeded without an authorized public identity." >&2 + exit 1 + fi + grep -q "SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON is required" store-package.log + test ! -d dist/chrome-web-store - name: Smoke direct server entrypoint run: python3 server/nano_server.py --help @@ -141,17 +149,6 @@ jobs: if-no-files-found: error retention-days: 30 - - name: Upload Chrome Web Store release candidate - if: matrix.node-version == '22.x' - uses: actions/upload-artifact@v7 - with: - name: selectpilot-chrome-web-store-${{ github.sha }} - path: | - dist/chrome-web-store/*.zip - dist/chrome-web-store/*.zip.json - if-no-files-found: error - retention-days: 30 - - name: Build entitlement authority image if: matrix.node-version == '22.x' run: docker build -f services/entitlement-authority/Dockerfile -t selectpilot-entitlement:${{ github.sha }} . diff --git a/.gitignore b/.gitignore index 6360eef..44ea49f 100644 --- a/.gitignore +++ b/.gitignore @@ -1,5 +1,6 @@ .DS_Store __pycache__/ +.venv/ node_modules/ package-lock.json npm-debug.log* diff --git a/README.md b/README.md index 715aa33..3a1986b 100644 --- a/README.md +++ b/README.md @@ -214,7 +214,7 @@ Store listing copy, privacy language, and the release checklist live in: - [`docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md`](docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md) - [`docs/PRIVACY_POLICY.md`](docs/PRIVACY_POLICY.md) -`pnpm package:store` builds and audits the upload ZIP. It fails closed unless the production entitlement public key is pinned and valid; unsigned or remotely executed commercial logic is never release-eligible. +`pnpm package:store` builds and audits the upload ZIP. The repository keyring is deliberately unprovisioned; an authorized release must supply `SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON`, which is validated and injected only into the isolated staging tree. The evidence report records the public key IDs and keyring digest. Missing identity, unsigned access, or remotely executed commercial logic blocks packaging. ## Compute Distribution diff --git a/ZERO_LEAKAGE.md b/ZERO_LEAKAGE.md index d87b8e8..f856891 100644 --- a/ZERO_LEAKAGE.md +++ b/ZERO_LEAKAGE.md @@ -9,7 +9,7 @@ The runtime is also profile-based: the product prefers the smallest viable local - Selected text is processed through a local bridge at `http://127.0.0.1:8083`. - The local bridge talks only to a locally running Ollama instance. - Cloud-hosted Ollama models are intentionally ignored for the core summarize, agent, and embed flows. -- If a requested model is unavailable locally, the bridge falls back only to other local models. +- If the exact policy-selected local model is unavailable, the request fails explicitly. A different local model may be used only when it is preinstalled, declared in signed runtime policy, non-quarantined, and surfaced to the user as a fallback; the current production policy declares no model fallback. - The first-run flow is expected to detect the runtime, install the small local profile, benchmark it, and assign the result before the extension is treated as ready. ## What runs locally diff --git a/agent/agent-client.js b/agent/agent-client.js index b3dc038..1725894 100644 --- a/agent/agent-client.js +++ b/agent/agent-client.js @@ -7,6 +7,10 @@ export async function runAgent(prompt, context) { return { reasoning: res.reasoning, markdown: res.markdown, - json: res.json ?? {} + json: res.json ?? {}, + model: res.model, + source: res.source, + routing: res.routing, + trace_id: res.trace_id, }; } diff --git a/agent/agent-client.ts b/agent/agent-client.ts index 0f6dccb..e74a219 100644 --- a/agent/agent-client.ts +++ b/agent/agent-client.ts @@ -9,6 +9,10 @@ export async function runAgent(prompt: string, context: AgentContext): Promise) ?? {} + json: (res.json as Record) ?? {}, + model: res.model, + source: res.source, + routing: res.routing, + trace_id: res.trace_id, }; } diff --git a/agent/agent-pipeline.js b/agent/agent-pipeline.js index 8d39ef7..c2712a4 100644 --- a/agent/agent-pipeline.js +++ b/agent/agent-pipeline.js @@ -53,6 +53,10 @@ export async function runPipeline(input, context, userPrompt) { return { reasoning: [...chain, ...(agentOut.reasoning || [])], markdown: agentOut.markdown, - json: agentOut.json + json: agentOut.json, + model: agentOut.model, + source: agentOut.source, + routing: agentOut.routing, + trace_id: agentOut.trace_id, }; } diff --git a/agent/agent-pipeline.ts b/agent/agent-pipeline.ts index a71727b..98bf8a2 100644 --- a/agent/agent-pipeline.ts +++ b/agent/agent-pipeline.ts @@ -62,6 +62,10 @@ export async function runPipeline(input: string, context: AgentContext, userProm return { reasoning: [...chain, ...(agentOut.reasoning || [])], markdown: agentOut.markdown, - json: agentOut.json + json: agentOut.json, + model: agentOut.model, + source: agentOut.source, + routing: agentOut.routing, + trace_id: agentOut.trace_id, }; } diff --git a/agent/agent-types.ts b/agent/agent-types.ts index dafa4ae..f2e99af 100644 --- a/agent/agent-types.ts +++ b/agent/agent-types.ts @@ -18,4 +18,12 @@ export type AgentOutput = { reasoning: string[]; markdown: string; json: Record; + model: string; + source: string; + routing: { + model: string; + num_ctx: number; + reason: string; + }; + trace_id?: string; }; diff --git a/api/nano-client.js b/api/nano-client.js index b67f2f3..3eb713a 100644 --- a/api/nano-client.js +++ b/api/nano-client.js @@ -3,16 +3,22 @@ import { endpoints } from './endpoints.js'; import { apiRequest } from './request.js'; export async function summarize(payload) { - return apiRequest(endpoints.summarize, { body: payload }); + return apiRequest(endpoints.summarize, { + body: payload, + timeoutMs: 45_000, + }); } export async function extract(payload) { - return apiRequest(endpoints.extract, { body: payload }); + return apiRequest(endpoints.extract, { body: payload, timeoutMs: 45_000 }); } export async function embed(payload) { return apiRequest(endpoints.embed, { body: payload }); } export async function agent(payload) { - return apiRequest(endpoints.agent, { body: payload }); + return apiRequest(endpoints.agent, { + body: payload, + timeoutMs: 45_000, + }); } export async function compileIntent(payload) { return apiRequest(endpoints.intentCompile, { body: payload }); diff --git a/api/nano-client.ts b/api/nano-client.ts index 10addfc..1433f9c 100644 --- a/api/nano-client.ts +++ b/api/nano-client.ts @@ -7,6 +7,17 @@ export type SummarizePayload = { text: string; url?: string; title?: string; met export type ExtractPayload = { text: string; preset?: string; url?: string; title?: string; metadata?: Record }; export type AgentPayload = { prompt: string; context?: Record }; export type EmbedPayload = { text: string }; +export type RuntimeRouting = { + model: string; + num_ctx: number; + reason: string; +}; +export type RuntimeResponseTruth = { + model: string; + source: string; + routing: RuntimeRouting; + trace_id?: string; +}; export type IntentCompilePayload = { intent: string; has_selection?: boolean; @@ -71,13 +82,22 @@ export type RuntimeMetaEvent = { }; export async function summarize(payload: SummarizePayload) { - return apiRequest<{ summary: string; markdown: string }>(endpoints.summarize, { body: payload }); + return apiRequest<{ summary: string; markdown: string } & RuntimeResponseTruth>(endpoints.summarize, { + body: payload, + timeoutMs: 45_000, + }); } export async function extract(payload: ExtractPayload) { - return apiRequest<{ preset: string; label: string; description: string; markdown: string; json: Record }>( + return apiRequest<{ + preset: string; + label: string; + description: string; + markdown: string; + json: Record; + } & RuntimeResponseTruth>( endpoints.extract, - { body: payload } + { body: payload, timeoutMs: 45_000 } ); } @@ -86,7 +106,10 @@ export async function embed(payload: EmbedPayload) { } export async function agent(payload: AgentPayload) { - return apiRequest<{ reasoning: string[]; markdown: string; json: unknown }>(endpoints.agent, { body: payload }); + return apiRequest<{ reasoning: string[]; markdown: string; json: unknown } & RuntimeResponseTruth>(endpoints.agent, { + body: payload, + timeoutMs: 45_000, + }); } export async function compileIntent(payload: IntentCompilePayload) { diff --git a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md index f2e2ea6..24d45a3 100644 --- a/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md +++ b/docs/CHROME_WEB_STORE_RELEASE_CHECKLIST.md @@ -15,7 +15,7 @@ Last updated: 2026-08-25 - [ ] `pnpm package:store` - [ ] Package SHA-256 and inventory are preserved with the release evidence. -`pnpm package:store` must fail if the pinned entitlement keyring is absent, invalid, or remote executable code enters the extension. Never bypass that gate with a development key or unsigned access. +`pricing/entitlement-public-keys.json` in the repository must remain unprovisioned. `pnpm package:store` must fail unless an authorized entitlement keyring is supplied explicitly, and it must also fail if that keyring is invalid or remote executable code enters the extension. Never bypass that gate with a development key or unsigned access. Production packaging requires `SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON` as a JSON key ring whose values are raw 32-byte Ed25519 public keys encoded as 64 hexadecimal characters. The matching issuer uses only: @@ -24,6 +24,8 @@ Production packaging requires `SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON` as a JS Never place the private key, its contents, or a production signature fixture in Git, CI logs, Linear, test data, or the extension package. Signer provisioning and activation remain an authorized secrets operation. +Run the dedicated `Chrome Web Store release candidate` workflow against the exact authorized revision. Its `chrome-web-store-production` environment must provide the keyring secret; ordinary pull-request CI verifies that packaging stays fail-closed and never uploads a production-looking ZIP. Packaging rejects a dirty checkout or a declared SHA that differs from `HEAD`. Reconcile the embedded source revision, artifact SHA-256, public key IDs, and keyring digest with the signer identity before Store upload. + The separately operated local billing authority installs its pinned runtime with `python3 -m pip install --requirement requirements-billing.txt`. Its wallet RPC is loopback-only, entitlement/order state is written with owner-only permissions, administrative revocation requires a configured secret of at least 32 characters, and generated order IDs contain 128 bits of random material. ## macOS Helper Release Gate diff --git a/docs/ENTITLEMENT_DEPLOYMENT.md b/docs/ENTITLEMENT_DEPLOYMENT.md index 5f0ba03..df90886 100644 --- a/docs/ENTITLEMENT_DEPLOYMENT.md +++ b/docs/ENTITLEMENT_DEPLOYMENT.md @@ -6,7 +6,7 @@ The authority is the only component allowed to issue SelectPilot access. It stor Create these files outside the repository with mode `0600`: -- Ed25519 PKCS#8 private key matching `pricing/entitlement-public-keys.json` +- Ed25519 PKCS#8 private key and its authorized public keyring, both provisioned outside the repository - Paddle webhook endpoint secret - Paddle price map JSON @@ -19,7 +19,7 @@ The price map shape is: } ``` -Copy `services/entitlement-authority/.env.example` to an external deployment `.env` and set the three corresponding `*_FILE` variables. The file contains paths, IDs, and ports only; secret values remain in the referenced mode-`0600` files. Create the state directory with mode `0700` and ownership matching `SELECTPILOT_UID`/`SELECTPILOT_GID`. +Copy `services/entitlement-authority/.env.example` to an external deployment `.env` and set the corresponding `*_FILE` variables. `SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE` is the external public identity used to prove that the private signer matches the Store package; the checked-in extension keyring remains empty. The environment file contains paths, IDs, and ports only; credential values remain in the referenced mode-`0600` files. Create the state directory with mode `0700` and ownership matching `SELECTPILOT_UID`/`SELECTPILOT_GID`. Run the fail-closed preflight before Compose: diff --git a/installer/macos/com.selectpilot.bridge.plist.template b/installer/macos/com.selectpilot.bridge.plist.template index e547e77..dad2fce 100644 --- a/installer/macos/com.selectpilot.bridge.plist.template +++ b/installer/macos/com.selectpilot.bridge.plist.template @@ -6,6 +6,7 @@ __PYTHON_BIN____PROJECT_ROOT__/server/nano_server.py --port8083 --run-dir__RUN_DIR__ + --state-dir__STATE_DIR__ --log-dir__LOG_DIR__ --binary-path__PROJECT_ROOT__/server/nano_server.py
    @@ -18,6 +19,7 @@ CHROMEAI_PORT_INFO__RUN_DIR__/port.info CHROMEAI_EXPECTED_PORT_FILE__RUN_DIR__/port.info CHROMEAI_ALLOWED_RANGE8083 + CHROMEAI_RUNTIME_STATE_DIR__STATE_DIR__ CHROMEAI_OLLAMA_BASE_URL__OLLAMA_BASE_URL__ CHROMEAI_RUNTIME_PROFILE__RUNTIME_PROFILE__ CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ @@ -29,6 +31,7 @@ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_AUTO_PULL_MODELS0 + CHROMEAI_MAX_CONCURRENT_OPERATIONS1 SELECTPILOT_ENTITLEMENT_AUTHORITY_URL__ENTITLEMENT_AUTHORITY_URL__
    diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall index 99ba646..bcdac6e 100755 --- a/installer/macos/scripts/postinstall +++ b/installer/macos/scripts/postinstall @@ -11,6 +11,7 @@ USER_HOME="$(dscl . -read "/Users/$CONSOLE_USER" NFSHomeDirectory | awk '{print USER_ID="$(id -u "$CONSOLE_USER")" SOURCE_ROOT="/Library/Application Support/SelectPilot" RUN_DIR="$USER_HOME/Library/Application Support/SelectPilot/run" +STATE_DIR="$USER_HOME/Library/Application Support/SelectPilot/state" LOG_DIR="$USER_HOME/Library/Logs/SelectPilot" AGENT_DIR="$USER_HOME/Library/LaunchAgents" AGENT_PATH="$AGENT_DIR/com.selectpilot.bridge.plist" @@ -47,11 +48,12 @@ IFS='|' read -r RUNTIME_PROFILE OLLAMA_MODEL OLLAMA_FAST_MODEL OLLAMA_EMBED_MODE $PROFILE_VALUES EOF -install -d -m 700 -o "$CONSOLE_USER" "$RUN_DIR" "$LOG_DIR" "$AGENT_DIR" +install -d -m 700 -o "$CONSOLE_USER" "$RUN_DIR" "$STATE_DIR" "$LOG_DIR" "$AGENT_DIR" sed \ -e "s|__PYTHON_BIN__|$PYTHON_BIN|g" \ -e "s|__PROJECT_ROOT__|$SOURCE_ROOT|g" \ -e "s|__RUN_DIR__|$RUN_DIR|g" \ + -e "s|__STATE_DIR__|$STATE_DIR|g" \ -e "s|__LOG_DIR__|$LOG_DIR|g" \ -e "s|__BINARY_HASH__|$HASH|g" \ -e "s|__OLLAMA_BASE_URL__|http://127.0.0.1:11434|g" \ diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index 6415fbe..f2b3610 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -9,6 +9,7 @@ __INSTALLED_BINARY__ --port8083 --run-dir__RUN_DIR__ + --state-dir__STATE_DIR__ --log-dir__LOG_DIR__ --binary-path__INSTALLED_BINARY__ @@ -22,6 +23,7 @@ CHROMEAI_PORT_INFO__RUN_DIR__/port.info CHROMEAI_EXPECTED_PORT_FILE__RUN_DIR__/port.info CHROMEAI_ALLOWED_RANGE8083 + CHROMEAI_RUNTIME_STATE_DIR__STATE_DIR__ CHROMEAI_OLLAMA_BASE_URL__OLLAMA_BASE_URL__ CHROMEAI_OLLAMA_MODEL__OLLAMA_MODEL__ CHROMEAI_OLLAMA_FAST_MODEL__OLLAMA_FAST_MODEL__ @@ -32,6 +34,7 @@ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_AUTO_PULL_MODELS0 + CHROMEAI_MAX_CONCURRENT_OPERATIONS1 diff --git a/package.json b/package.json index 6fadbe2..2c7b90e 100644 --- a/package.json +++ b/package.json @@ -42,7 +42,7 @@ "test:privacy": "node --test tests/panel/network-boundary.test.mjs && python3 -m unittest tests/server/test_privacy_proof.py", "test:billing": "python3 -m unittest tests/server/test_billing_security.py tests/server/test_entitlement_signer.py && python3 -c \"import sys; sys.path.insert(0, 'server'); import monero_payment_server\"", "typecheck": "tsc --noEmit -p tsconfig.json", - "validate:server": "python3 server/nano_server.py --run-dir /tmp/chromeai-run --log-dir /tmp/chromeai-log" + "validate:server": "python3 server/nano_server.py --run-dir /tmp/chromeai-run --state-dir /tmp/chromeai-state --log-dir /tmp/chromeai-log" }, "devDependencies": { "@eslint/js": "^10.0.1", diff --git a/panel/panel.css b/panel/panel.css index 7b5f244..49c1c1c 100644 --- a/panel/panel.css +++ b/panel/panel.css @@ -131,7 +131,7 @@ body[data-reduce-motion="true"] *::after { animation: none !important; transitio .truth-strip { display: grid; - grid-template-columns: repeat(3, minmax(0, 1fr)); + grid-template-columns: repeat(auto-fit, minmax(112px, 1fr)); gap: 8px; } diff --git a/panel/panel.html b/panel/panel.html index e4507e4..36b7c20 100644 --- a/panel/panel.html +++ b/panel/panel.html @@ -23,6 +23,10 @@ Execution Detecting +
    + Model + Detecting +
    Boundary Local-only @@ -32,6 +36,10 @@ Verifying Awaiting proof
    +
    + Profile + Detecting +
    diff --git a/panel/panel.js b/panel/panel.js index 07da0c5..c65c7bb 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -102,6 +102,7 @@ let memorySnapshot = { lastUpdatedAt: null, }; let entitlementSnapshot = null; +let entitlementStatusKnown = false; let bridgeAvailable = false; let installationSnapshot = { state: 'idle', @@ -146,6 +147,37 @@ function setStatusBar(text) { if (statusBar) statusBar.textContent = text; } +function applyResponseRuntimeTruth(taskFamily, response) { + const value = response; + const routing = value?.routing; + if (typeof value?.model !== 'string' + || !value.model + || typeof value?.source !== 'string' + || !value.source + || typeof routing?.model !== 'string' + || routing.model !== value.model + || typeof routing.num_ctx !== 'number' + || !Number.isInteger(routing.num_ctx) + || Number(routing.num_ctx) <= 0 + || typeof routing.reason !== 'string' + || !routing.reason) { + throw new Error('Runtime response omitted its exact local model contract'); + } + applyRuntimeEvent({ + taskFamily, + selectedModel: routing.model, + executionGeography: 'local', + }); + const profile = getEffectiveRecommendedProfile(); + const reason = routing.reason.replace(/_/g, ' '); + if (truthExecutionEl) + truthExecutionEl.textContent = 'Local'; + if (truthModelEl) + truthModelEl.textContent = routing.model; + if (truthProfileEl) + truthProfileEl.textContent = `${profile.label} · ${Number(routing.num_ctx).toLocaleString()} ctx`; + setStatusBar(`${taskFamily} · ${routing.model} · ${Number(routing.num_ctx).toLocaleString()} ctx · ${reason}`); +} function setLeakageFeedback(status, details) { if (leakageStatusEl) leakageStatusEl.textContent = status; @@ -747,6 +779,10 @@ function renderMemoryState() { function renderEntitlementStatus() { if (!entitlementStatusEl) return; + if (!entitlementStatusKnown) { + entitlementStatusEl.textContent = 'Access state unavailable · refresh required.'; + return; + } const tier = entitlementSnapshot?.tier || 'essential'; const token = entitlementSnapshot?.token; if (!token) { @@ -760,9 +796,11 @@ function renderEntitlementStatus() { async function refreshEntitlementStatus() { try { entitlementSnapshot = await request('entitlement:get'); + entitlementStatusKnown = true; } catch { entitlementSnapshot = null; + entitlementStatusKnown = false; } renderEntitlementStatus(); renderSelectionState(); @@ -1037,23 +1075,20 @@ async function refreshTier() { async function refreshRuntime() { const startedAt = performance.now(); try { - try { - runtimeProfilesPayload = await fetchRuntimeProfiles(); - } - catch { - runtimeProfilesPayload = { - profiles: RUNTIME_PROFILES, - recommended_profile: 'fast', - reason: 'The smallest viable profile is the safest starting point.', - }; - } + runtimeProfilesPayload = await fetchRuntimeProfiles(); const health = await fetchHealth(); bridgeAvailable = true; try { await refreshInstallationStatus(); } catch { - installationSnapshot = { state: 'idle', label: 'Ready to install', progress: 0, profile: null, action_required: null }; + installationSnapshot = { + state: 'error', + label: 'Installation state unavailable', + progress: 0, + profile: null, + action_required: 'Refresh the local bridge before changing runtime state.', + }; } privacyProofSnapshot = await fetchPrivacyProof(); runtimeSnapshot = { @@ -1140,15 +1175,7 @@ async function refreshRuntime() { async function doSummarize() { setStatus('Summarizing selected text...'); const res = await request('panel:summarize'); - if (res?.model_selection) { - applyRuntimeEvent({ - taskFamily: 'summarize', - selectedModel: String(res.model_selection.model || res.model || 'unknown'), - selectionPath: res.model_selection.selection_path, - policyVersion: res.model_selection.policy_version ?? null, - executionGeography: 'local', - }); - } + applyResponseRuntimeTruth('summarize', res); renderOutput({ title: 'Summary', markdown: res.markdown || res.summary || '', @@ -1163,15 +1190,7 @@ async function doExtract(presetKey) { const selectedPreset = getExtractionPreset(presetKey || extractPresetEl?.value); setStatus(`Extracting ${selectedPreset.label.toLowerCase()}...`); const res = await request('panel:extract', { preset: selectedPreset.key }); - if (res?.model_selection) { - applyRuntimeEvent({ - taskFamily: 'extract', - selectedModel: String(res.model_selection.model || res.model || 'unknown'), - selectionPath: res.model_selection.selection_path, - policyVersion: res.model_selection.policy_version ?? null, - executionGeography: 'local', - }); - } + applyResponseRuntimeTruth('extract', res); renderOutput({ title: res.label || selectedPreset.label, markdown: res.markdown || '', @@ -1186,6 +1205,7 @@ async function doExtract(presetKey) { async function doFirstRunExample() { const selectedPreset = getExtractionPreset(FIRST_RUN_EXAMPLE.preset); const res = await request('panel:extract_demo'); + applyResponseRuntimeTruth('extract', res); renderOutput({ title: res.label || selectedPreset.label, markdown: res.markdown || '', @@ -1204,6 +1224,7 @@ async function doRewrite() { const prompt = agentPromptEl?.value.trim() || 'Rewrite the selected text in clearer, tighter language.'; setStatus('Rewriting...'); const res = await request('panel:agent', { prompt }); + applyResponseRuntimeTruth('agent', res); const markdown = res.markdown || ''; renderOutput({ title: 'Rewrite', @@ -1223,15 +1244,7 @@ async function doAsk() { const prompt = agentPromptEl?.value.trim() || 'Answer the question using the selected text as context.'; setStatus('Asking Ollama...'); const res = await request('panel:agent', { prompt }); - if (res?.model_selection) { - applyRuntimeEvent({ - taskFamily: 'agent', - selectedModel: String(res.model_selection.model || res.model || 'unknown'), - selectionPath: res.model_selection.selection_path, - policyVersion: res.model_selection.policy_version ?? null, - executionGeography: 'local', - }); - } + applyResponseRuntimeTruth('agent', res); const markdown = res.markdown || ''; renderOutput({ title: 'Answer', diff --git a/panel/panel.ts b/panel/panel.ts index d5b0c79..969263f 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -196,6 +196,7 @@ let memorySnapshot: MemoryStatusSnapshot = { lastUpdatedAt: null, }; let entitlementSnapshot: EntitlementSnapshot = null; +let entitlementStatusKnown = false; let bridgeAvailable = false; let installationSnapshot = { state: 'idle', @@ -242,6 +243,42 @@ function setStatusBar(text: string) { if (statusBar) statusBar.textContent = text; } +function applyResponseRuntimeTruth(taskFamily: 'extract' | 'summarize' | 'agent', response: unknown) { + const value = response as { + model?: unknown; + source?: unknown; + routing?: { model?: unknown; num_ctx?: unknown; reason?: unknown }; + }; + const routing = value?.routing; + if ( + typeof value?.model !== 'string' + || !value.model + || typeof value?.source !== 'string' + || !value.source + || typeof routing?.model !== 'string' + || routing.model !== value.model + || typeof routing.num_ctx !== 'number' + || !Number.isInteger(routing.num_ctx) + || Number(routing.num_ctx) <= 0 + || typeof routing.reason !== 'string' + || !routing.reason + ) { + throw new Error('Runtime response omitted its exact local model contract'); + } + + applyRuntimeEvent({ + taskFamily, + selectedModel: routing.model, + executionGeography: 'local', + }); + const profile = getEffectiveRecommendedProfile(); + const reason = routing.reason.replace(/_/g, ' '); + if (truthExecutionEl) truthExecutionEl.textContent = 'Local'; + if (truthModelEl) truthModelEl.textContent = routing.model; + if (truthProfileEl) truthProfileEl.textContent = `${profile.label} · ${Number(routing.num_ctx).toLocaleString()} ctx`; + setStatusBar(`${taskFamily} · ${routing.model} · ${Number(routing.num_ctx).toLocaleString()} ctx · ${reason}`); +} + function setLeakageFeedback(status: string, details: string) { if (leakageStatusEl) leakageStatusEl.textContent = status; if (leakageDetailsEl) leakageDetailsEl.textContent = details; @@ -898,6 +935,10 @@ function renderMemoryState() { function renderEntitlementStatus() { if (!entitlementStatusEl) return; + if (!entitlementStatusKnown) { + entitlementStatusEl.textContent = 'Access state unavailable · refresh required.'; + return; + } const tier = entitlementSnapshot?.tier || 'essential'; const token = entitlementSnapshot?.token; if (!token) { @@ -912,8 +953,10 @@ function renderEntitlementStatus() { async function refreshEntitlementStatus() { try { entitlementSnapshot = await request('entitlement:get'); + entitlementStatusKnown = true; } catch { entitlementSnapshot = null; + entitlementStatusKnown = false; } renderEntitlementStatus(); renderSelectionState(); @@ -1192,21 +1235,19 @@ async function refreshTier() { async function refreshRuntime() { const startedAt = performance.now(); try { - try { - runtimeProfilesPayload = await fetchRuntimeProfiles(); - } catch { - runtimeProfilesPayload = { - profiles: RUNTIME_PROFILES, - recommended_profile: 'fast', - reason: 'The smallest viable profile is the safest starting point.', - }; - } + runtimeProfilesPayload = await fetchRuntimeProfiles(); const health = await fetchHealth(); bridgeAvailable = true; try { await refreshInstallationStatus(); } catch { - installationSnapshot = { state: 'idle', label: 'Ready to install', progress: 0, profile: null, action_required: null }; + installationSnapshot = { + state: 'error', + label: 'Installation state unavailable', + progress: 0, + profile: null, + action_required: 'Refresh the local bridge before changing runtime state.', + }; } privacyProofSnapshot = await fetchPrivacyProof(); runtimeSnapshot = { @@ -1285,15 +1326,7 @@ async function refreshRuntime() { async function doSummarize() { setStatus('Summarizing selected text...'); const res = await request('panel:summarize'); - if (res?.model_selection) { - applyRuntimeEvent({ - taskFamily: 'summarize', - selectedModel: String(res.model_selection.model || res.model || 'unknown'), - selectionPath: res.model_selection.selection_path, - policyVersion: res.model_selection.policy_version ?? null, - executionGeography: 'local', - }); - } + applyResponseRuntimeTruth('summarize', res); renderOutput({ title: 'Summary', markdown: res.markdown || res.summary || '', @@ -1309,15 +1342,7 @@ async function doExtract(presetKey?: ExtractionPresetKey) { const selectedPreset = getExtractionPreset(presetKey || extractPresetEl?.value); setStatus(`Extracting ${selectedPreset.label.toLowerCase()}...`); const res = await request('panel:extract', { preset: selectedPreset.key }); - if (res?.model_selection) { - applyRuntimeEvent({ - taskFamily: 'extract', - selectedModel: String(res.model_selection.model || res.model || 'unknown'), - selectionPath: res.model_selection.selection_path, - policyVersion: res.model_selection.policy_version ?? null, - executionGeography: 'local', - }); - } + applyResponseRuntimeTruth('extract', res); renderOutput({ title: res.label || selectedPreset.label, markdown: res.markdown || '', @@ -1333,6 +1358,7 @@ async function doExtract(presetKey?: ExtractionPresetKey) { async function doFirstRunExample() { const selectedPreset = getExtractionPreset(FIRST_RUN_EXAMPLE.preset); const res = await request('panel:extract_demo'); + applyResponseRuntimeTruth('extract', res); renderOutput({ title: res.label || selectedPreset.label, markdown: res.markdown || '', @@ -1352,6 +1378,7 @@ async function doRewrite() { const prompt = agentPromptEl?.value.trim() || 'Rewrite the selected text in clearer, tighter language.'; setStatus('Rewriting...'); const res = await request('panel:agent', { prompt }); + applyResponseRuntimeTruth('agent', res); const markdown = res.markdown || ''; renderOutput({ title: 'Rewrite', @@ -1373,15 +1400,7 @@ async function doAsk() { const prompt = agentPromptEl?.value.trim() || 'Answer the question using the selected text as context.'; setStatus('Asking Ollama...'); const res = await request('panel:agent', { prompt }); - if (res?.model_selection) { - applyRuntimeEvent({ - taskFamily: 'agent', - selectedModel: String(res.model_selection.model || res.model || 'unknown'), - selectionPath: res.model_selection.selection_path, - policyVersion: res.model_selection.policy_version ?? null, - executionGeography: 'local', - }); - } + applyResponseRuntimeTruth('agent', res); const markdown = res.markdown || ''; renderOutput({ title: 'Answer', diff --git a/pricing/entitlement-public-keys.json b/pricing/entitlement-public-keys.json index 0ff0fca..c6e1b77 100644 --- a/pricing/entitlement-public-keys.json +++ b/pricing/entitlement-public-keys.json @@ -1,11 +1,4 @@ { "schema_version": 1, - "keys": [ - { - "kid": "selectpilot-prod-2026-01", - "alg": "Ed25519", - "public_key_hex": "e9c1429a51ccf7d5eac11c4d22139a85414231d3ad64578e1fdb7f004ed74bd7", - "status": "active" - } - ] + "keys": [] } diff --git a/scripts/build-extension.mjs b/scripts/build-extension.mjs index d6f41b5..dba6bb6 100644 --- a/scripts/build-extension.mjs +++ b/scripts/build-extension.mjs @@ -3,37 +3,11 @@ // spec_ref: "reporting" import { build } from 'esbuild'; -import { readFile, writeFile } from 'node:fs/promises'; import path from 'node:path'; import { fileURLToPath } from 'node:url'; const projectRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); -async function injectEntitlementPublicKeys() { - const raw = process.env.SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON; - if (!raw) return; - let keys; - try { - keys = JSON.parse(raw); - } catch { - throw new Error('SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON must be valid JSON'); - } - const entries = Object.entries(keys ?? {}); - if (entries.length === 0 || entries.some(([kid, key]) => !/^[A-Za-z0-9._-]{1,64}$/.test(kid) || !/^[0-9a-f]{64}$/i.test(String(key)))) { - throw new Error('Entitlement public key ring must contain valid key IDs and 32-byte Ed25519 public keys'); - } - const target = path.join(projectRoot, 'background/entitlement-service.js'); - const source = await readFile(target, 'utf8'); - const configured = source.replace( - /const ENTITLEMENT_PUBLIC_KEYS = \{[\s\S]*?\n\};/, - `const ENTITLEMENT_PUBLIC_KEYS = ${JSON.stringify(keys)};`, - ); - if (configured === source) throw new Error('Entitlement public key injection marker not found'); - await writeFile(target, configured); -} - -await injectEntitlementPublicKeys(); - await build({ entryPoints: [path.join(projectRoot, 'content/content-script.ts')], outfile: path.join(projectRoot, 'content/content-script.bundle.js'), diff --git a/scripts/entitlement-keyring.mjs b/scripts/entitlement-keyring.mjs new file mode 100644 index 0000000..9733e4a --- /dev/null +++ b/scripts/entitlement-keyring.mjs @@ -0,0 +1,78 @@ +// module_name: entitlement_keyring_contract +// spec_ref: "security" + +const KEY_ID_PATTERN = /^[A-Za-z0-9._-]{1,64}$/; +const PUBLIC_KEY_PATTERN = /^[0-9a-f]{64}$/i; + +function assertKeyEntry(key) { + if ( + !key + || typeof key !== 'object' + || !KEY_ID_PATTERN.test(String(key.kid || '')) + || key.alg !== 'Ed25519' + || !PUBLIC_KEY_PATTERN.test(String(key.public_key_hex || '')) + || !['active', 'retiring'].includes(key.status) + ) { + throw new Error('Entitlement public keyring contains an invalid Ed25519 key entry'); + } +} + +// @spec_ref security +export function validateEntitlementKeyring(value, { requireActive = true } = {}) { + if (!value || typeof value !== 'object' || value.schema_version !== 1 || !Array.isArray(value.keys)) { + throw new Error('Entitlement public keyring must use schema_version 1 with a keys array'); + } + + value.keys.forEach(assertKeyEntry); + const ids = value.keys.map((key) => key.kid); + if (new Set(ids).size !== ids.length) throw new Error('Entitlement public key IDs must be unique'); + + const active = value.keys.filter((key) => key.status === 'active'); + if (requireActive && active.length !== 1) { + throw new Error('Entitlement public keyring must contain exactly one active signing identity'); + } + if (!requireActive && active.length > 0) { + throw new Error('Repository entitlement keyring must remain unprovisioned'); + } + + return { + schema_version: 1, + keys: value.keys.map((key) => ({ + kid: key.kid, + alg: 'Ed25519', + public_key_hex: key.public_key_hex.toLowerCase(), + status: key.status, + })), + }; +} + +// @spec_ref security +export function parseProvisionedEntitlementKeyring(raw) { + if (typeof raw !== 'string' || !raw.trim()) { + throw new Error('Store package blocked: SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON is required'); + } + + let parsed; + try { + parsed = JSON.parse(raw); + } catch { + throw new Error('SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON must be valid JSON'); + } + + if (parsed?.schema_version === 1 && Array.isArray(parsed.keys)) { + return validateEntitlementKeyring(parsed); + } + + const entries = parsed && typeof parsed === 'object' && !Array.isArray(parsed) + ? Object.entries(parsed) + : []; + if (entries.length !== 1) { + throw new Error('Map-form entitlement keyring must contain exactly one active key'); + } + + const [kid, publicKeyHex] = entries[0]; + return validateEntitlementKeyring({ + schema_version: 1, + keys: [{ kid, alg: 'Ed25519', public_key_hex: String(publicKeyHex), status: 'active' }], + }); +} diff --git a/scripts/install-entitlement-systemd.sh b/scripts/install-entitlement-systemd.sh index 8216564..1249282 100755 --- a/scripts/install-entitlement-systemd.sh +++ b/scripts/install-entitlement-systemd.sh @@ -13,6 +13,7 @@ UNIT_TARGET=/etc/systemd/system/selectpilot-entitlement.service for secret_file in \ /etc/selectpilot/secrets/entitlement-ed25519.pem \ + /etc/selectpilot/secrets/entitlement-public-keys.json \ /etc/selectpilot/secrets/paddle-webhook-secret \ /etc/selectpilot/secrets/paddle-price-map.json do @@ -32,6 +33,7 @@ if ss -ltn | awk '{print $4}' | grep -Eq '(^|:)8091$'; then fi SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE=/etc/selectpilot/secrets/entitlement-ed25519.pem \ +SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE=/etc/selectpilot/secrets/entitlement-public-keys.json \ node "$ROOT/scripts/verify-entitlement-key.mjs" install -d -o root -g root -m 0755 "$APP_DIR" diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index 4c1e110..cef570b 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -22,6 +22,7 @@ OLLAMA_FAST_NUM_CTX="${CHROMEAI_OLLAMA_FAST_NUM_CTX:-$OLLAMA_NUM_CTX}" MAX_INPUT_CHARS="${CHROMEAI_MAX_INPUT_CHARS:-16000}" OLLAMA_SEED="${CHROMEAI_OLLAMA_SEED:-42}" RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run}" +STATE_DIR="${CHROMEAI_RUNTIME_STATE_DIR:-${HOME}/Library/Application Support/SelectPilot/state}" LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}" PYTHON_BIN="${CHROMEAI_PYTHON_BIN:-$(command -v python3)}" @@ -31,7 +32,7 @@ if ! "$PYTHON_BIN" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, fi mkdir -p "${HOME}/Library/LaunchAgents" "$INSTALL_DIR" "$APP_DIR/presets" "$APP_DIR/runtime" -mkdir -p "$RUN_DIR" "$LOG_DIR" +mkdir -p "$RUN_DIR" "$STATE_DIR" "$LOG_DIR" # LaunchAgents cannot reliably read repositories located in macOS privacy- # protected folders such as Documents. Install an immutable runtime copy in @@ -50,6 +51,7 @@ sed \ -e "s|__INSTALL_DIR__|$INSTALL_DIR|g" \ -e "s|__INSTALLED_BINARY__|$INSTALLED_BINARY|g" \ -e "s|__RUN_DIR__|$RUN_DIR|g" \ + -e "s|__STATE_DIR__|$STATE_DIR|g" \ -e "s|__LOG_DIR__|$LOG_DIR|g" \ -e "s|__BINARY_HASH__|$HASH|g" \ -e "s|__OLLAMA_BASE_URL__|$OLLAMA_BASE_URL|g" \ @@ -81,6 +83,7 @@ Next steps: 6. Ollama context window: $OLLAMA_NUM_CTX 7. Deterministic seed: $OLLAMA_SEED 8. Run dir: $RUN_DIR - 9. Log dir: $LOG_DIR - 10. Run 'pnpm benchmark:local' to validate latency on this machine. + 9. Runtime state dir: $STATE_DIR + 10. Log dir: $LOG_DIR + 11. Run 'pnpm benchmark:local' to validate latency on this machine. EOF diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs index e089a6c..d9ee977 100644 --- a/scripts/package-chrome-store.mjs +++ b/scripts/package-chrome-store.mjs @@ -7,6 +7,7 @@ import { cp, mkdir, readFile, readdir, rm, stat, utimes, writeFile } from 'node: import path from 'node:path'; import { spawnSync } from 'node:child_process'; import { fileURLToPath } from 'node:url'; +import { parseProvisionedEntitlementKeyring, validateEntitlementKeyring } from './entitlement-keyring.mjs'; import { validateStoreAssets } from './validate-store-assets.mjs'; const projectRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); @@ -64,20 +65,7 @@ export async function assertReleaseSafe(files, root = projectRoot) { throw new Error('Store package blocked: production entitlement signature verification is not configured (SOD-837).'); } const keyring = JSON.parse(await readFile(path.join(root, 'pricing/entitlement-public-keys.json'), 'utf8')); - if ( - keyring.schema_version !== 1 - || !Array.isArray(keyring.keys) - || !keyring.keys.some((key) => key.status === 'active') - || keyring.keys.some((key) => ( - typeof key.kid !== 'string' - || !key.kid - || key.alg !== 'Ed25519' - || !/^[0-9a-f]{64}$/i.test(key.public_key_hex || '') - || !['active', 'retiring'].includes(key.status) - )) - ) { - throw new Error('Store package blocked: production entitlement keyring is invalid.'); - } + validateEntitlementKeyring(keyring); if (files.some((file) => file.startsWith('billing/') || file === 'pricing/paddle-products.json')) { throw new Error('Store package blocked: inactive remote checkout code or product placeholders entered runtime inventory.'); } @@ -98,20 +86,52 @@ export async function assertReleaseSafe(files, root = projectRoot) { } } +export async function assertSourceKeyringUnprovisioned(root = projectRoot) { + const keyring = JSON.parse(await readFile(path.join(root, 'pricing/entitlement-public-keys.json'), 'utf8')); + validateEntitlementKeyring(keyring, { requireActive: false }); +} + async function sha256(filePath) { return createHash('sha256').update(await readFile(filePath)).digest('hex'); } -export async function packageChromeStore(root = projectRoot) { - await validateStoreAssets(root); - const files = await collectRuntimeFiles(root); - await assertReleaseSafe(files, root); +function resolveSourceRevision(root, declaredRevision) { + const revisionResult = spawnSync('git', ['rev-parse', 'HEAD'], { cwd: root, encoding: 'utf8' }); + const actualRevision = revisionResult.status === 0 ? revisionResult.stdout.trim() : ''; + const revision = declaredRevision || actualRevision; + if (!/^[0-9a-f]{40}$/i.test(revision)) { + throw new Error('Store package blocked: an exact 40-character source revision is required'); + } + if (actualRevision && revision.toLowerCase() !== actualRevision.toLowerCase()) { + throw new Error('Store package blocked: declared source revision does not match checked-out HEAD'); + } + if (actualRevision) { + const status = spawnSync('git', ['status', '--porcelain', '--untracked-files=no'], { cwd: root, encoding: 'utf8' }); + if (status.status !== 0 || status.stdout.trim()) { + throw new Error('Store package blocked: tracked source is dirty after build'); + } + } + return revision.toLowerCase(); +} +export async function packageChromeStore( + root = projectRoot, + { + entitlementKeyringJson = process.env.SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON, + sourceRevision = process.env.SELECTPILOT_SOURCE_SHA, + } = {}, +) { const packageJson = JSON.parse(await readFile(path.join(root, 'package.json'), 'utf8')); const outputRoot = path.join(root, 'dist', 'chrome-web-store'); const stageRoot = path.join(outputRoot, `selectpilot-${packageJson.version}`); const zipPath = `${stageRoot}.zip`; await rm(outputRoot, { recursive: true, force: true }); + + await validateStoreAssets(root); + await assertSourceKeyringUnprovisioned(root); + const provisionedKeyring = parseProvisionedEntitlementKeyring(entitlementKeyringJson); + const exactSourceRevision = resolveSourceRevision(root, sourceRevision); + const files = await collectRuntimeFiles(root); await mkdir(stageRoot, { recursive: true }); for (const relative of files) { @@ -121,6 +141,12 @@ export async function packageChromeStore(root = projectRoot) { await utimes(destination, fixedDate, fixedDate); } + const stagedKeyringPath = path.join(stageRoot, 'pricing/entitlement-public-keys.json'); + const canonicalKeyring = `${JSON.stringify(provisionedKeyring, null, 2)}\n`; + await writeFile(stagedKeyringPath, canonicalKeyring); + await utimes(stagedKeyringPath, fixedDate, fixedDate); + await assertReleaseSafe(files, stageRoot); + const zip = spawnSync('zip', ['-X', '-q', zipPath, ...files], { cwd: stageRoot, encoding: 'utf8', @@ -131,8 +157,11 @@ export async function packageChromeStore(root = projectRoot) { const report = { schema_version: 1, version: packageJson.version, + source_revision: exactSourceRevision, artifact: path.basename(zipPath), sha256: await sha256(zipPath), + entitlement_keyring_sha256: createHash('sha256').update(canonicalKeyring).digest('hex'), + entitlement_key_ids: provisionedKeyring.keys.map((key) => key.kid), file_count: files.length, files, }; diff --git a/scripts/preflight-entitlement-deployment.sh b/scripts/preflight-entitlement-deployment.sh index f25560d..0a51e26 100755 --- a/scripts/preflight-entitlement-deployment.sh +++ b/scripts/preflight-entitlement-deployment.sh @@ -23,6 +23,7 @@ for name in \ SELECTPILOT_AUTHORITY_HOST_PORT \ SELECTPILOT_ENTITLEMENT_KEY_ID \ SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE \ + SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE \ SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE \ SELECTPILOT_PADDLE_PRICE_MAP_FILE \ SELECTPILOT_STATE_DIR @@ -56,6 +57,7 @@ file_owner() { for secret_file in \ "$SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE" \ + "$SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE" \ "$SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE" \ "$SELECTPILOT_PADDLE_PRICE_MAP_FILE" do @@ -85,6 +87,7 @@ for (const [id, entry] of Object.entries(map)) { NODE SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE="$SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE" \ +SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE="$SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE" \ node "$ROOT/scripts/verify-entitlement-key.mjs" docker compose --env-file "$ENV_FILE" -f "$COMPOSE_FILE" config --quiet diff --git a/scripts/spec-compiler.mjs b/scripts/spec-compiler.mjs index 863c352..af89fa6 100644 --- a/scripts/spec-compiler.mjs +++ b/scripts/spec-compiler.mjs @@ -35,7 +35,7 @@ const REQUIRED_TOP_LEVEL_NODES = [ ]; const SCAN_EXTENSIONS = new Set(['.ts', '.tsx', '.js', '.mjs', '.py']); -const SCAN_IGNORE = new Set(['.git', 'node_modules', 'test-results', 'dist', 'site-dist']); +const SCAN_IGNORE = new Set(['.git', '.venv', 'node_modules', 'test-results', 'dist', 'site-dist']); const SIMULATION_EVIDENCE = { evidence_class: 'deterministic_simulation', runtime_verified: false, diff --git a/scripts/verify-entitlement-key.mjs b/scripts/verify-entitlement-key.mjs index 328f8dd..6c27782 100644 --- a/scripts/verify-entitlement-key.mjs +++ b/scripts/verify-entitlement-key.mjs @@ -2,10 +2,14 @@ // spec_ref: "validation_layer" import { createPrivateKey, createPublicKey, sign, verify } from 'node:crypto'; import { readFile } from 'node:fs/promises'; +import { parseProvisionedEntitlementKeyring } from './entitlement-keyring.mjs'; const privateKeyFile = process.env.SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE; if (!privateKeyFile) throw new Error('SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE is required'); -const keyring = JSON.parse(await readFile('pricing/entitlement-public-keys.json', 'utf8')); +const publicKeyringFile = process.env.SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE; +const publicKeyringJson = process.env.SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON + || (publicKeyringFile ? await readFile(publicKeyringFile, 'utf8') : ''); +const keyring = parseProvisionedEntitlementKeyring(publicKeyringJson); const active = keyring.keys.filter((key) => key.status === 'active'); if (active.length !== 1) throw new Error('Exactly one active entitlement key is required'); const privateKey = createPrivateKey(await readFile(privateKeyFile)); diff --git a/server/nano_server.py b/server/nano_server.py index fbce321..13d35aa 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -13,6 +13,7 @@ import os import re import socket +import sys import threading import time from collections import deque @@ -33,6 +34,7 @@ LOCAL_HOSTS = {"127.0.0.1", "localhost", "::1"} MAX_REQUEST_BYTES = 2 * 1024 * 1024 CHROME_EXTENSION_ORIGIN = re.compile(r"^chrome-extension://[a-p]{32}$") +INFERENCE_ENDPOINTS = {"/summarize", "/extract", "/agent", "/embed", "/benchmark"} ALLOWED_BRIDGE_ENDPOINT_PATHS = [ "/health", @@ -135,6 +137,49 @@ def __init__(self, code: str, message: str, *, status: int = 400, details: dict self.details = details or {} +class RuntimeStateError(RuntimeError): + """Raised when decision-affecting local runtime state cannot be read safely.""" + + +class LocalOperationAdmission: + def __init__(self, limit: int): + if limit < 1 or limit > 4: + raise ValueError("local operation concurrency must be between 1 and 4") + self.limit = limit + self._semaphore = threading.BoundedSemaphore(limit) + self._lock = threading.Lock() + self._active = 0 + + def try_acquire(self) -> bool: + if not self._semaphore.acquire(blocking=False): + return False + with self._lock: + self._active += 1 + return True + + def release(self) -> None: + with self._lock: + if self._active < 1: + raise RuntimeError("local operation admission release imbalance") + self._active -= 1 + self._semaphore.release() + + def snapshot(self) -> dict: + with self._lock: + return {"active": self._active, "limit": self.limit} + + +def configured_operation_concurrency() -> int: + raw = os.environ.get("CHROMEAI_MAX_CONCURRENT_OPERATIONS", "1") + try: + return int(raw) + except ValueError as exc: + raise RuntimeError("invalid_local_operation_concurrency") from exc + + +OPERATION_ADMISSION = LocalOperationAdmission(configured_operation_concurrency()) + + def extension_origin_allowed(origin: str | None) -> bool: candidate = str(origin or "").strip() if not candidate: @@ -445,6 +490,13 @@ def verify_binary(path: Path, expected_hash: str | None = None) -> bool: print(f"binary hash mismatch: {digest} != {expected_hash}") return False return True + + +def require_binary_integrity(path: Path, expected_hash: str | None = None) -> None: + if not verify_binary(path, expected_hash): + raise RuntimeError("runtime_integrity_check_failed") + + def ensure_dirs(path: Path): path.mkdir(parents=True, exist_ok=True) @@ -458,7 +510,9 @@ def write_port_info(port_file: Path, port: int): RUNTIME_DIR = PROJECT_ROOT / "runtime" RUNTIME_POLICY_PATH = RUNTIME_DIR / "model_policy.json" RUNTIME_REGISTRY_PATH = RUNTIME_DIR / "model_registry.runtime.json" -LIVE_FEEDBACK_PATH = RUNTIME_DIR / "live_feedback.jsonl" +DEFAULT_RUNTIME_STATE_DIR = Path(os.path.expanduser('~/Library/Application Support/SelectPilot/state')) +RUNTIME_STATE_DIR = Path(os.environ.get('CHROMEAI_RUNTIME_STATE_DIR', DEFAULT_RUNTIME_STATE_DIR)) +LIVE_FEEDBACK_PATH = RUNTIME_STATE_DIR / "live_feedback.jsonl" _LIVE_FEEDBACK_LOCK = threading.Lock() MAX_LIVE_FEEDBACK_BYTES = 2 * 1024 * 1024 MAX_RUNTIME_META_STREAMS = 4 @@ -488,9 +542,34 @@ def write_port_info(port_file: Path, port: int): }, ] -def _append_live_feedback(event: dict) -> None: +def configure_runtime_state(state_dir: Path) -> None: + global RUNTIME_STATE_DIR, LIVE_FEEDBACK_PATH + resolved = state_dir.expanduser().resolve() + ensure_dirs(resolved) + probe = resolved / f".write-probe-{os.getpid()}" try: - ensure_dirs(RUNTIME_DIR) + probe.write_text("selectpilot-runtime-state\n", encoding="utf-8") + probe.unlink() + except OSError as exc: + raise RuntimeError("runtime_state_directory_unwritable") from exc + RUNTIME_STATE_DIR = resolved + LIVE_FEEDBACK_PATH = resolved / "live_feedback.jsonl" + + +def _report_runtime_state_error(operation: str, exc: Exception) -> None: + payload = { + "event": "runtime_state.failed", + "operation": operation, + "status": "error", + "error": type(exc).__name__, + "ts": datetime.now(timezone.utc).isoformat(), + } + print(json.dumps(payload, ensure_ascii=False, separators=(",", ":")), file=sys.stderr) + + +def _append_live_feedback(event: dict) -> bool: + try: + ensure_dirs(LIVE_FEEDBACK_PATH.parent) entry = { "timestamp": datetime.now(timezone.utc).isoformat(), **event, @@ -502,9 +581,10 @@ def _append_live_feedback(event: dict) -> None: os.replace(LIVE_FEEDBACK_PATH, rotated) with LIVE_FEEDBACK_PATH.open("a", encoding="utf-8") as handle: handle.write(f"{line}\n") - except Exception: - # Feedback sink must not break request handling. - return + return True + except (OSError, TypeError, ValueError) as exc: + _report_runtime_state_error("write_feedback", exc) + return False def _load_json_file(path: Path) -> dict | None: @@ -649,21 +729,21 @@ def _recent_model_feedback(model_id: str, limit: int = MODEL_FAILURE_ISOLATION_W for line in reversed(lines): try: item = json.loads(line) - except json.JSONDecodeError: - continue + except json.JSONDecodeError as exc: + raise RuntimeStateError("runtime_state_feedback_corrupt") from exc if item.get("type") == "model_feedback" and item.get("model_id") == model_id: records.append(item) if len(records) >= limit: break - except OSError: - return [] + except OSError as exc: + raise RuntimeStateError("runtime_state_feedback_unreadable") from exc return list(reversed(records)) -def record_model_feedback(model_id: str, *, success: bool, retries: int, latency_ms: int, cancelled: bool = False) -> None: +def record_model_feedback(model_id: str, *, success: bool, retries: int, latency_ms: int, cancelled: bool = False) -> bool: if not model_id: - return - _append_live_feedback({ + return True + return _append_live_feedback({ "type": "model_feedback", "model_id": model_id, "success": bool(success), @@ -1178,15 +1258,9 @@ def select_smallest_sufficient_model(task_analysis: dict) -> dict: def fixed_temperature_for_task(task_type: str) -> float: - values = { - "classification": 0.0, - "extract": 0.0, - "rewrite": 0.3, - "analyze": 0.2, - "summarize": 0.2, - "agent": 0.2, - } - return float(values.get(str(task_type), 0.2)) + # Every admitted production generation route is deterministic. This + # metadata must describe the payload actually sent by OllamaClient. + return 0.0 def run_with_output_enforcement( @@ -1429,6 +1503,15 @@ def license_claim(payload: dict) -> dict: class Handler(BaseHTTPRequestHandler): server_version = "ChromeAINano/1.0" + _operation_slot_acquired = False + + def finish(self): + try: + super().finish() + finally: + if self._operation_slot_acquired: + OPERATION_ADMISSION.release() + self._operation_slot_acquired = False def _set_headers(self): self.send_header("Content-Type", "application/json") @@ -1477,6 +1560,7 @@ def do_GET(self): "ok": bool(health.get("reachable")) and bool(health.get("model_available")), "service": self.server_version, "ollama": health, + "local_operation_capacity": OPERATION_ADMISSION.snapshot(), }) return if path == '/privacy-proof': @@ -1579,6 +1663,16 @@ def do_POST(self): if contract is None: self._write_json(404, {"error": "not_found"}) return + if path in INFERENCE_ENDPOINTS: + if not OPERATION_ADMISSION.try_acquire(): + self._write_error( + 429, + "local_runtime_busy", + "Local runtime capacity is in use; retry after the active operation completes", + OPERATION_ADMISSION.snapshot(), + ) + return + self._operation_slot_acquired = True header_map = {k.lower(): v for k, v in self.headers.items()} trace_id = _resolve_trace_id(payload, header_map) @@ -1821,6 +1915,20 @@ def _execute_summarize(): ) self._write_error(e.status, e.code, e.message, e.details) return + except RuntimeStateError as e: + duration_ms = round((time.perf_counter() - started_at) * 1000) + _report_runtime_state_error("read_feedback", e) + emit_runtime_meta( + event_type="RUNTIME_FAILED", + trace_id=trace_id, + operation=contract.name, + status="error", + message="Local runtime state is unavailable", + duration_ms=duration_ms, + details={"code": str(e)}, + ) + self._write_error(503, str(e), "Local runtime state is unavailable") + return except (RuntimeError, OllamaError) as e: duration_ms = round((time.perf_counter() - started_at) * 1000) if runtime_model_id and path in {'/summarize', '/agent', '/extract'}: @@ -1890,11 +1998,17 @@ def _execute_summarize(): duration_ms = round((time.perf_counter() - started_at) * 1000) + state_persisted = True if runtime_model_id and path in {'/summarize', '/agent', '/extract'}: - record_model_feedback(runtime_model_id, success=True, retries=retry_count, latency_ms=duration_ms) + state_persisted = record_model_feedback( + runtime_model_id, + success=True, + retries=retry_count, + latency_ms=duration_ms, + ) model_selection = resp.get("model_selection") if isinstance(resp, dict) else None - _append_live_feedback({ + state_persisted = _append_live_feedback({ "trace_id": trace_id, "operation": contract.name, "endpoint": contract.endpoint, @@ -1917,7 +2031,20 @@ def _execute_summarize(): if isinstance(model_selection, dict) else None ), - }) + }) and state_persisted + + if not state_persisted: + emit_runtime_meta( + event_type="RUNTIME_FAILED", + trace_id=trace_id, + operation=contract.name, + status="error", + message="Local runtime state could not be persisted", + duration_ms=duration_ms, + details={"code": "runtime_state_write_failed"}, + ) + self._write_error(503, "runtime_state_write_failed", "Local runtime state could not be persisted") + return log_runtime_event( "request.completed", @@ -1944,20 +2071,21 @@ def main(): parser.add_argument('--port-range', default=None) parser.add_argument('--run-dir', default=os.path.expanduser('~/Library/Application Support/SelectPilot/run')) parser.add_argument('--log-dir', default=os.path.expanduser('~/Library/Logs/SelectPilot')) + parser.add_argument('--state-dir', default=os.environ.get('CHROMEAI_RUNTIME_STATE_DIR', str(DEFAULT_RUNTIME_STATE_DIR))) parser.add_argument('--binary-path', default=None) parser.add_argument('--binary-hash', default=None) args = parser.parse_args() run_dir = Path(args.run_dir) log_dir = Path(args.log_dir) - ensure_dirs(RUNTIME_DIR) + configure_runtime_state(Path(args.state_dir)) ensure_dirs(run_dir) ensure_dirs(log_dir) port_file = run_dir / 'port.info' binary_path = Path(args.binary_path) if args.binary_path else Path(__file__) expected = args.binary_hash or os.environ.get('CHROMEAI_BINARY_HASH') - verify_binary(binary_path, expected) + require_binary_integrity(binary_path, expected) auto_pull = os.environ.get('CHROMEAI_AUTO_PULL_MODELS', '0').strip().lower() not in {'0', 'false', 'no'} if auto_pull: diff --git a/services/entitlement-authority/.env.example b/services/entitlement-authority/.env.example index 345ea23..3c2b1d0 100644 --- a/services/entitlement-authority/.env.example +++ b/services/entitlement-authority/.env.example @@ -3,6 +3,7 @@ SELECTPILOT_GID=1000 SELECTPILOT_AUTHORITY_HOST_PORT=8091 SELECTPILOT_ENTITLEMENT_KEY_ID=selectpilot-prod-2026-01 SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE=/etc/selectpilot/secrets/entitlement-ed25519.pem +SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_FILE=/etc/selectpilot/secrets/entitlement-public-keys.json SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE=/etc/selectpilot/secrets/paddle-webhook-secret SELECTPILOT_PADDLE_PRICE_MAP_FILE=/etc/selectpilot/secrets/paddle-price-map.json SELECTPILOT_STATE_DIR=/var/lib/selectpilot-authority diff --git a/tests/e2e/extension-first-run-flow.spec.mjs b/tests/e2e/extension-first-run-flow.spec.mjs index 60209ae..fb45932 100644 --- a/tests/e2e/extension-first-run-flow.spec.mjs +++ b/tests/e2e/extension-first-run-flow.spec.mjs @@ -108,6 +108,13 @@ test('real extension preserves privacy from selected text through rendered outpu if (path === '/extract') { extractRequests.push(request.postDataJSON()); return fulfill({ + model: 'gemma4:e2b-it-qat', + source: 'ollama', + routing: { + model: 'gemma4:e2b-it-qat', + num_ctx: 16_384, + reason: 'smallest_qualified_structured_model', + }, preset: 'action_brief', label: 'Action Brief', description: 'Turn selected text into a concise action-oriented brief.', @@ -156,6 +163,9 @@ test('real extension preserves privacy from selected text through rendered outpu }); await expect(panelPage.locator('#result-title')).toHaveText('Action Brief'); + await expect(panelPage.locator('#truth-model')).toHaveText('gemma4:e2b-it-qat'); + await expect(panelPage.locator('#truth-profile')).toContainText('16,384 ctx'); + await expect(panelPage.locator('#status-bar')).toContainText('smallest qualified structured model'); await expect(panelPage.locator('#exports')).toContainText('Copy Markdown'); await panelPage.locator('#tab-structured').click(); await expect(panelPage.locator('#workflow')).toContainText('Publishing waits for both checks'); diff --git a/tests/e2e/panel-harness.html b/tests/e2e/panel-harness.html index ba699de..2c53eae 100644 --- a/tests/e2e/panel-harness.html +++ b/tests/e2e/panel-harness.html @@ -210,6 +210,9 @@

    Ready

    } if (msg.type === 'panel:extract_demo') { return { + model: 'gemma4:e2b-it-qat', + source: 'ollama', + routing: { model: 'gemma4:e2b-it-qat', num_ctx: 16384, reason: 'smallest_qualified_structured_model' }, label: 'Action Brief', description: 'Turn selected text into a concise action-oriented brief.', markdown: '## Action Brief\n\n- Maya: privacy check by Thursday\n- Jonas: verify payment and store assets', @@ -225,6 +228,9 @@

    Ready

    if (msg.type === 'panel:extract') { pushMemory('extract'); return { + model: 'gemma4:e2b-it-qat', + source: 'ollama', + routing: { model: 'gemma4:e2b-it-qat', num_ctx: 16384, reason: 'smallest_qualified_structured_model' }, label: 'Action Brief', description: 'Extract action items', markdown: '## Action Brief\n\n- Ship update\n- Inform team', @@ -235,7 +241,13 @@

    Ready

    }, }; } - return { markdown: 'ok', json: {} }; + return { + model: 'gemma4:e2b-it-qat', + source: 'ollama', + routing: { model: 'gemma4:e2b-it-qat', num_ctx: 16384, reason: 'qualified_general_model' }, + markdown: 'ok', + json: {}, + }; }, }, tabs: { diff --git a/tests/e2e/playwright.config.mjs b/tests/e2e/playwright.config.mjs index ffb4794..58279a5 100644 --- a/tests/e2e/playwright.config.mjs +++ b/tests/e2e/playwright.config.mjs @@ -14,7 +14,7 @@ export default defineConfig({ workers: process.env.CI ? 1 : undefined, reporter: [['list']], webServer: { - command: 'python3 server/nano_server.py --run-dir /tmp/selectpilot-e2e-run --log-dir /tmp/selectpilot-e2e-log', + command: 'python3 server/nano_server.py --run-dir /tmp/selectpilot-e2e-run --state-dir /tmp/selectpilot-e2e-state --log-dir /tmp/selectpilot-e2e-log', cwd: projectRoot, env: { ...process.env, diff --git a/tests/e2e/privacy-and-runtime.spec.mjs b/tests/e2e/privacy-and-runtime.spec.mjs index 6a2d1a0..b2016aa 100644 --- a/tests/e2e/privacy-and-runtime.spec.mjs +++ b/tests/e2e/privacy-and-runtime.spec.mjs @@ -25,6 +25,7 @@ test('health endpoint is reachable and exposes ollama boundary state', async ({ const health = await res.json(); expect(health).toHaveProperty('ollama'); expect(health.ollama).toHaveProperty('privacy_mode'); + expect(health.local_operation_capacity).toEqual({ active: 0, limit: 1 }); }); test('mocked selected-text extraction response shape', async () => { @@ -43,4 +44,4 @@ test('mocked selected-text extraction response shape', async () => { expect(mock.label).toBe('Action Brief'); expect(mock.markdown).toContain('Action Brief'); expect(mock.json.task).toBeTruthy(); -}); \ No newline at end of file +}); diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index 80e2e33..fe0dcc7 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -82,7 +82,11 @@ test('production macOS package also binds launchd to Python 3.9 or newer', () => assert.match(postinstall, /sys\.version_info >= \(3, 9\)/); assert.match(postinstall, /s\|__PYTHON_BIN__\|\$PYTHON_BIN\|g/); + assert.match(postinstall, /s\|__STATE_DIR__\|\$STATE_DIR\|g/); assert.match(launchAgent, /__PYTHON_BIN__/); + assert.match(launchAgent, /--state-dir/); + assert.match(launchAgent, /CHROMEAI_RUNTIME_STATE_DIR/); + assert.match(launchAgent, /CHROMEAI_MAX_CONCURRENT_OPERATIONS<\/key>1/); assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); }); diff --git a/tests/panel/capability-truth.test.mjs b/tests/panel/capability-truth.test.mjs index 5c594e5..6468945 100644 --- a/tests/panel/capability-truth.test.mjs +++ b/tests/panel/capability-truth.test.mjs @@ -18,4 +18,6 @@ test('unimplemented media capabilities are absent from the shipped extension', a assert.doesNotMatch(panel, /btn-(?:transcribe|vision)/); assert.doesNotMatch(harness, /btn-(?:transcribe|vision)/); assert.doesNotMatch(background, /panel:(?:transcribe|vision)/); + assert.match(panel, /id="truth-model"/); + assert.match(panel, /id="truth-profile"/); }); diff --git a/tests/panel/store-package.test.mjs b/tests/panel/store-package.test.mjs index 4e461a1..5abef06 100644 --- a/tests/panel/store-package.test.mjs +++ b/tests/panel/store-package.test.mjs @@ -4,11 +4,16 @@ import assert from 'node:assert/strict'; import { execFileSync, spawnSync } from 'node:child_process'; import test from 'node:test'; -import { mkdir, mkdtemp, readFile, writeFile } from 'node:fs/promises'; +import { cp, mkdir, mkdtemp, readFile, writeFile } from 'node:fs/promises'; import { tmpdir } from 'node:os'; import path from 'node:path'; import { fileURLToPath } from 'node:url'; -import { collectRuntimeFiles, assertReleaseSafe } from '../../scripts/package-chrome-store.mjs'; +import { + collectRuntimeFiles, + assertReleaseSafe, + assertSourceKeyringUnprovisioned, + packageChromeStore, +} from '../../scripts/package-chrome-store.mjs'; import { REQUIRED_IMAGES, readPngDimensions, validateStoreAssets } from '../../scripts/validate-store-assets.mjs'; const root = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '../..'); @@ -37,9 +42,8 @@ test('runtime inventory excludes source, tests, reports, and transient files', a assert.ok(files.every((file) => !/(^|\/)(tests|reports|node_modules)(\/|$)/.test(file))); }); -test('store release accepts the pinned production entitlement keyring', async () => { - const files = await collectRuntimeFiles(root); - await assert.doesNotReject(assertReleaseSafe(files, root)); +test('repository keyring remains deliberately unprovisioned', async () => { + await assert.doesNotReject(assertSourceKeyringUnprovisioned(root)); }); test('store runtime captures selected text only', async () => { @@ -54,7 +58,7 @@ test('store runtime captures selected text only', async () => { assert.ok(Object.values(tiers).flat().every((feature) => !/(?:audio|video|image|multimodal)/i.test(feature))); }); -test('relative CLI paths execute validation and deterministic packaging', () => { +test('relative CLI paths validate assets and block packaging without explicit release identity', () => { const validation = execFileSync(process.execPath, ['./scripts/validate-store-assets.mjs'], { cwd: root, encoding: 'utf8', @@ -64,9 +68,37 @@ test('relative CLI paths execute validation and deterministic packaging', () => const packaging = spawnSync(process.execPath, ['./scripts/package-chrome-store.mjs'], { cwd: root, encoding: 'utf8', + env: { ...process.env, SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON: '' }, }); - assert.equal(packaging.status, 0, packaging.stderr); - assert.match(packaging.stdout, /^selectpilot-1\.0\.0\.zip [0-9a-f]{64}\s*$/); + assert.notEqual(packaging.status, 0); + assert.match(packaging.stderr, /SELECTPILOT_ENTITLEMENT_PUBLIC_KEYS_JSON is required/); +}); + +test('explicit public identity is injected only into the isolated Store staging tree', async () => { + const releaseRoot = await mkdtemp(path.join(tmpdir(), 'selectpilot-store-package-')); + const files = await collectRuntimeFiles(root); + for (const relative of new Set([...files, ...REQUIRED_IMAGES.keys(), 'package.json'])) { + const destination = path.join(releaseRoot, relative); + await mkdir(path.dirname(destination), { recursive: true }); + await cp(path.join(root, relative), destination); + } + + const publicKey = 'ab'.repeat(32); + const report = await packageChromeStore(releaseRoot, { + entitlementKeyringJson: JSON.stringify({ 'selectpilot-test-release': publicKey }), + sourceRevision: '1'.repeat(40), + }); + const stagedKeyring = JSON.parse(await readFile( + path.join(releaseRoot, 'dist/chrome-web-store/selectpilot-1.0.0/pricing/entitlement-public-keys.json'), + 'utf8', + )); + const sourceKeyring = JSON.parse(await readFile(path.join(releaseRoot, 'pricing/entitlement-public-keys.json'), 'utf8')); + + assert.deepEqual(report.entitlement_key_ids, ['selectpilot-test-release']); + assert.equal(report.source_revision, '1'.repeat(40)); + assert.match(report.entitlement_keyring_sha256, /^[0-9a-f]{64}$/); + assert.equal(stagedKeyring.keys[0].public_key_hex, publicKey); + assert.deepEqual(sourceKeyring.keys, []); }); test('store release rejects remotely hosted executable code', async () => { diff --git a/tests/server/test_capability_truth.py b/tests/server/test_capability_truth.py index 79570ec..a4edc79 100644 --- a/tests/server/test_capability_truth.py +++ b/tests/server/test_capability_truth.py @@ -1,8 +1,13 @@ # module_name: tests_server_test_capability_truth_py # spec_ref: "testing_strategy.integration_tests" +import io import sys +import tempfile import unittest +from contextlib import redirect_stdout +from hashlib import sha256 from pathlib import Path +from unittest.mock import patch SERVER_DIR = Path(__file__).resolve().parents[2] / "server" @@ -18,6 +23,51 @@ def test_unimplemented_media_routes_are_not_admitted(self): self.assertNotIn("/transcribe", nano_server.OPERATION_CONTRACTS) self.assertNotIn("/vision", nano_server.OPERATION_CONTRACTS) + def test_binary_integrity_mismatch_stops_startup_contract(self): + with tempfile.TemporaryDirectory() as temp_dir: + binary = Path(temp_dir) / "nano_server.py" + binary.write_text("print('trusted')\n", encoding="utf-8") + expected = sha256(binary.read_bytes()).hexdigest() + nano_server.require_binary_integrity(binary, expected) + + binary.write_text("print('tampered')\n", encoding="utf-8") + with redirect_stdout(io.StringIO()): + with self.assertRaisesRegex(RuntimeError, "runtime_integrity_check_failed"): + nano_server.require_binary_integrity(binary, expected) + + def test_runtime_feedback_uses_explicit_mutable_state_directory(self): + with tempfile.TemporaryDirectory() as temp_dir: + state_dir = Path(temp_dir) / "state" + with patch.object(nano_server, "RUNTIME_STATE_DIR", nano_server.RUNTIME_STATE_DIR), patch.object( + nano_server, "LIVE_FEEDBACK_PATH", nano_server.LIVE_FEEDBACK_PATH + ): + nano_server.configure_runtime_state(state_dir) + self.assertTrue(nano_server._append_live_feedback({"type": "test"})) + self.assertEqual(nano_server.LIVE_FEEDBACK_PATH.parent, state_dir.resolve()) + self.assertTrue(nano_server.LIVE_FEEDBACK_PATH.is_file()) + + def test_runtime_feedback_write_failure_is_explicit(self): + with tempfile.TemporaryDirectory() as temp_dir: + feedback_path = Path(temp_dir) / "feedback.jsonl" + with patch.object(nano_server, "LIVE_FEEDBACK_PATH", feedback_path), patch.object( + Path, "open", side_effect=OSError("read only") + ), patch("sys.stderr") as stderr: + self.assertFalse(nano_server._append_live_feedback({"type": "test"})) + self.assertTrue(stderr.write.called) + + def test_production_intent_metadata_matches_zero_temperature_runtime(self): + for task in ("classification", "extract", "rewrite", "analyze", "summarize", "agent"): + self.assertEqual(nano_server.fixed_temperature_for_task(task), 0.0) + + def test_local_operation_admission_is_bounded_and_recoverable(self): + admission = nano_server.LocalOperationAdmission(1) + self.assertTrue(admission.try_acquire()) + self.assertFalse(admission.try_acquire()) + self.assertEqual(admission.snapshot(), {"active": 1, "limit": 1}) + admission.release() + self.assertTrue(admission.try_acquire()) + admission.release() + if __name__ == "__main__": unittest.main() From 93334b73e58bc122947eb6b8b866f3a2e65e0e50 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 00:12:05 +0200 Subject: [PATCH 58/66] docs(governance): record release hardening evidence Record the verified repository boundary and preserve every external production and review gate as open. Verified: evidence commands recorded in the checkpoint. --- .../20-checkpoint.md | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md index bfc4264..4fc7804 100644 --- a/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md +++ b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md @@ -2,21 +2,21 @@ ## TodoCheckpointDraft -- Current todo: inventory full architecture, contracts, tests, threat model, abuse surfaces, and negative space at exact head `937e4926944052495c8fc626f2d36cb422fbec4d`. -- Completed: project identity verified; branch fast-forwarded; global constitution read; signed lifecycle policy cryptographically verified; signed intent created; Store flow paused. -- Active slice: read-only gap and abuse-surface audit. -- Evidence: PR #22 remains open and review-blocked; signed Axiom intent `AXIOM-20260830-OXBH25`; lifecycle policy hash `932bbb231cb5e1aeb35d6bbe6327db6f3b1b3cf1e125e38366cbccbbc59434cf`. -- Blocked-on: installed NovaForge operator release references a missing `config/axiom-lifecycle-policy-binding.json`; production authorities remain external but do not block repository hardening. -- Next: enumerate existing controls and tests, run fresh remote baseline, then classify and repair material findings. +- Current todo: publish the signed repository-hardening candidate to PR #22, observe exact-head GitHub admission, and retain the external production gates without laundering them into code completion. +- Completed: Linear global runtime laws, SelectPilot project truth, and GitHub PR state were re-read before mutation; the full-system gap/blind-spot/negative-space audit was refreshed; the signed intent remains valid. +- Repository repairs: Store packaging now requires an explicit external entitlement public identity and exact clean SHA; ordinary CI proves the missing-identity failure path; mutable helper state is user-owned; binary mismatch stops startup; local inference concurrency is bounded; exact model/context routing survives the agent pipeline and is visible in the extension. +- Remote evidence: Hetzner passed lint, typecheck/build parity, 50 Node tests, 73 server tests, 7 Playwright journeys, privacy, billing, manifest/assets, negative Store packaging, frontier traceability, runtime-policy validation, stress contracts, and exact-SHA code admission. The admission report correctly remained `runtime_verified=false`, `physical_hardware_verified=false`, and `full_system_ok=false`. +- External truth: PR #22 still requires one eligible independent approval; production DNS/TLS, Paddle credentials/lifecycle, authorized entitlement signer, Developer ID Installer/notarization, exact Store upload/publication, and physical Apple first-use evidence remain outside repository authority. +- Next: push the candidate into PR #22, require exact-head GitHub checks and independent review, then execute the external gates only when their real authorities exist. ## ResumeStateHint -Resume from this checkpoint, re-read `10-intent.md`, verify `intent.json`, compare current head/worktree with this file, and refresh PR/Linear state before any completion claim. +Resume from this checkpoint, re-read `10-intent.md`, verify `intent.json`, compare current head/worktree with this file, and refresh PR/Linear/external authority state before any completion claim. ## DriftCheckDraft - Intent alignment: yes. - Scope alignment: yes. - New fallback or duplicate owner: no. -- Evidence sufficiency: not yet; audit and stress proof pending. -- Decision: continue. +- Evidence sufficiency: repository hardening is locally and remotely qualified; exact-head GitHub, physical runtime, protected review, and production activation evidence are still required. +- Decision: continue through GitHub admission; fail closed on every external completion claim until its authority is present. From 80643126f4dabff0c6ddd6235698a110626f6322 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 00:39:02 +0200 Subject: [PATCH 59/66] fix(runtime): bind helper integrity and timeout failures Cover active helper modules and runtime policy with an aggregate launch-time hash, install the missing installation manager dependency, and normalize raw request timeouts into the explicit helper failure contract. Verified: targeted Python runtime tests, bootstrap contract tests, shell syntax checks, and git diff --check. --- README.md | 2 + .../20-checkpoint.md | 17 +++++---- .../com.selectpilot.bridge.plist.template | 1 + installer/macos/scripts/postinstall | 12 +++++- launchd/com.chromeai.nano.plist | 1 + scripts/install-macos-local.sh | 11 ++++++ server/nano_server.py | 38 +++++++++++++++++++ server/ollama_client.py | 4 ++ tests/panel/bootstrap-contract.test.mjs | 4 ++ tests/server/test_bridge_origin.py | 25 +++++++++++- tests/server/test_capability_truth.py | 34 +++++++++++++++++ tests/server/test_ollama_client.py | 21 ++++++++++ 12 files changed, 159 insertions(+), 11 deletions(-) diff --git a/README.md b/README.md index 3a1986b..999ad43 100644 --- a/README.md +++ b/README.md @@ -69,6 +69,8 @@ curl -sSf http://127.0.0.1:8083/health Browser requests to the bridge are accepted only from a Chrome extension origin. Release operators should set `SELECTPILOT_EXTENSION_ORIGIN=chrome-extension://` to pin the exact installed identity; ordinary local command-line health checks omit `Origin` and remain available. +The macOS installers bind the LaunchAgent to a deterministic `CHROMEAI_RUNTIME_HASH` covering the active bridge modules, extraction preset, model policy, runtime registry, and promotion audit. Any installed-file drift stops startup with `runtime_integrity_check_failed`; reinstall the exact authorized candidate instead of editing the installed runtime in place. + --- ## Architecture diff --git a/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md index 4fc7804..c30b207 100644 --- a/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md +++ b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md @@ -2,12 +2,13 @@ ## TodoCheckpointDraft -- Current todo: publish the signed repository-hardening candidate to PR #22, observe exact-head GitHub admission, and retain the external production gates without laundering them into code completion. -- Completed: Linear global runtime laws, SelectPilot project truth, and GitHub PR state were re-read before mutation; the full-system gap/blind-spot/negative-space audit was refreshed; the signed intent remains valid. -- Repository repairs: Store packaging now requires an explicit external entitlement public identity and exact clean SHA; ordinary CI proves the missing-identity failure path; mutable helper state is user-owned; binary mismatch stops startup; local inference concurrency is bounded; exact model/context routing survives the agent pipeline and is visible in the extension. -- Remote evidence: Hetzner passed lint, typecheck/build parity, 50 Node tests, 73 server tests, 7 Playwright journeys, privacy, billing, manifest/assets, negative Store packaging, frontier traceability, runtime-policy validation, stress contracts, and exact-SHA code admission. The admission report correctly remained `runtime_verified=false`, `physical_hardware_verified=false`, and `full_system_ok=false`. -- External truth: PR #22 still requires one eligible independent approval; production DNS/TLS, Paddle credentials/lifecycle, authorized entitlement signer, Developer ID Installer/notarization, exact Store upload/publication, and physical Apple first-use evidence remain outside repository authority. -- Next: push the candidate into PR #22, require exact-head GitHub checks and independent review, then execute the external gates only when their real authorities exist. +- Current todo: qualify the physical-smoke follow-up on Hetzner and GitHub, preserve the independent-review gate, and retain every unavailable production authority without laundering it into code completion. +- Completed: Linear global runtime laws, SelectPilot project truth, and GitHub PR state were re-read before mutation; the full-system gap/blind-spot/negative-space audit was refreshed; the signed intent remains valid. PR #22 source `93334b73e58bc122947eb6b8b866f3a2e65e0e50` passed all 9 exact-head GitHub checks in CI run `33812111260` and CodeQL run `33812111254`. +- Repository repairs: Store packaging now requires an explicit external entitlement public identity and exact clean SHA; ordinary CI proves the missing-identity failure path; mutable helper state is user-owned; local inference concurrency is bounded; exact model/context routing survives the agent pipeline and is visible in the extension. Physical inspection also found and repaired two installer/runtime gaps: transport `TimeoutError` now becomes explicit bridge HTTP 503, and both local and packaged LaunchAgents bind an aggregate hash over all active runtime code and policy. The local installer now includes the previously omitted `installation_manager.py` dependency. +- Remote evidence: Hetzner and exact-head GitHub passed lint, typecheck/build parity, 50 Node tests, 73 server tests, 7 Playwright journeys, privacy, billing, manifest/assets, negative Store packaging, frontier traceability, runtime-policy validation, stress contracts, and exact-SHA code admission. The admission report correctly remained `runtime_verified=false`, `physical_hardware_verified=false`, and `full_system_ok=false`. +- Physical M4 evidence: the explicit local prewarm completed in `130.599s` (`129.834s` model load) from external model storage. Once warm, two identical `action_brief` extractions completed in `3.722s` and `3.134s`, selected `gemma4:e2b-it-qat` at 16,384 context with `smallest_qualified_structured_model`, and produced identical canonical digest `8283d9866499bef776bc4d171b94d772944a04330ffed329774baa438c55b6d4`. An explicitly missing model returned degraded health and HTTP 503 without substitution; restoring the exact model recovered extraction in `3.446s`. +- External truth: PR #22 still requires one eligible independent approval; production DNS/TLS, Paddle credentials/lifecycle, authorized entitlement signer, Developer ID Installer/notarization, exact Store upload/publication, full browser first-use/export evidence, and remaining declared Apple hardware profiles remain outside repository authority. +- Next: obtain full remote regression and a new signed exact-head GitHub result for the physical-smoke repair, then execute protected merge and production gates only when their real authorities exist. ## ResumeStateHint @@ -18,5 +19,5 @@ Resume from this checkpoint, re-read `10-intent.md`, verify `intent.json`, compa - Intent alignment: yes. - Scope alignment: yes. - New fallback or duplicate owner: no. -- Evidence sufficiency: repository hardening is locally and remotely qualified; exact-head GitHub, physical runtime, protected review, and production activation evidence are still required. -- Decision: continue through GitHub admission; fail closed on every external completion claim until its authority is present. +- Evidence sufficiency: repository hardening and one physical M4 structured-runtime slice are qualified; final follow-up GitHub admission, complete physical/browser runtime, protected review, and production activation evidence are still required. +- Decision: continue through remote and GitHub admission; fail closed on every full-system completion claim until its authority is present. diff --git a/installer/macos/com.selectpilot.bridge.plist.template b/installer/macos/com.selectpilot.bridge.plist.template index dad2fce..11069e8 100644 --- a/installer/macos/com.selectpilot.bridge.plist.template +++ b/installer/macos/com.selectpilot.bridge.plist.template @@ -30,6 +30,7 @@ CHROMEAI_MAX_INPUT_CHARS__MAX_INPUT_CHARS__ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ + CHROMEAI_RUNTIME_HASH__RUNTIME_HASH__ CHROMEAI_AUTO_PULL_MODELS0 CHROMEAI_MAX_CONCURRENT_OPERATIONS1 SELECTPILOT_ENTITLEMENT_AUTHORITY_URL__ENTITLEMENT_AUTHORITY_URL__ diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall index bcdac6e..1164281 100755 --- a/installer/macos/scripts/postinstall +++ b/installer/macos/scripts/postinstall @@ -15,7 +15,6 @@ STATE_DIR="$USER_HOME/Library/Application Support/SelectPilot/state" LOG_DIR="$USER_HOME/Library/Logs/SelectPilot" AGENT_DIR="$USER_HOME/Library/LaunchAgents" AGENT_PATH="$AGENT_DIR/com.selectpilot.bridge.plist" -HASH="$(shasum -a 256 "$SOURCE_ROOT/server/nano_server.py" | awk '{print $1}')" PYTHON_BIN="" for candidate in /opt/homebrew/bin/python3 /usr/local/bin/python3 /Library/Frameworks/Python.framework/Versions/Current/bin/python3 /usr/bin/python3; do if [ -x "$candidate" ] && "$candidate" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, 9) else 1)'; then @@ -27,6 +26,16 @@ if [ -z "$PYTHON_BIN" ]; then echo "SelectPilot requires Python 3.9 or newer." >&2 exit 1 fi +HASH="$(shasum -a 256 "$SOURCE_ROOT/server/nano_server.py" | awk '{print $1}')" +CHROMEAI_RUNTIME_HASH="$(PYTHONPATH="$SOURCE_ROOT/server" "$PYTHON_BIN" - "$SOURCE_ROOT" <<'PY' +import sys +from pathlib import Path + +from nano_server import runtime_integrity_digest + +print(runtime_integrity_digest(Path(sys.argv[1]))) +PY +)" PROFILE_VALUES="$(sudo -u "$CONSOLE_USER" "$PYTHON_BIN" - "$SOURCE_ROOT/server" <<'PY' import sys @@ -56,6 +65,7 @@ sed \ -e "s|__STATE_DIR__|$STATE_DIR|g" \ -e "s|__LOG_DIR__|$LOG_DIR|g" \ -e "s|__BINARY_HASH__|$HASH|g" \ + -e "s|__RUNTIME_HASH__|$CHROMEAI_RUNTIME_HASH|g" \ -e "s|__OLLAMA_BASE_URL__|http://127.0.0.1:11434|g" \ -e "s|__RUNTIME_PROFILE__|$RUNTIME_PROFILE|g" \ -e "s|__OLLAMA_MODEL__|$OLLAMA_MODEL|g" \ diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index f2b3610..df86fcc 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -33,6 +33,7 @@ CHROMEAI_MAX_INPUT_CHARS__MAX_INPUT_CHARS__ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ CHROMEAI_BINARY_HASH__BINARY_HASH__ + CHROMEAI_RUNTIME_HASH__RUNTIME_HASH__ CHROMEAI_AUTO_PULL_MODELS0 CHROMEAI_MAX_CONCURRENT_OPERATIONS1 diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index cef570b..6e315bf 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -12,6 +12,7 @@ RUNTIME_MODULES=( ollama_client.py extraction_presets.py runtime_profiles.py + installation_manager.py ) OLLAMA_BASE_URL="${CHROMEAI_OLLAMA_BASE_URL:-http://127.0.0.1:11434}" OLLAMA_MODEL="${CHROMEAI_OLLAMA_MODEL:-gemma4:e2b-it-qat}" @@ -45,6 +46,15 @@ for policy in model_policy.json model_registry.runtime.json promotion_audit.json install -m 0444 "$ROOT/runtime/$policy" "$APP_DIR/runtime/$policy" done HASH="$(shasum -a 256 "$INSTALLED_BINARY" | awk '{print $1}')" +CHROMEAI_RUNTIME_HASH="$(PYTHONPATH="$INSTALL_DIR" "$PYTHON_BIN" - "$APP_DIR" <<'PY' +import sys +from pathlib import Path + +from nano_server import runtime_integrity_digest + +print(runtime_integrity_digest(Path(sys.argv[1]))) +PY +)" sed \ -e "s|__PYTHON_BIN__|$PYTHON_BIN|g" \ @@ -54,6 +64,7 @@ sed \ -e "s|__STATE_DIR__|$STATE_DIR|g" \ -e "s|__LOG_DIR__|$LOG_DIR|g" \ -e "s|__BINARY_HASH__|$HASH|g" \ + -e "s|__RUNTIME_HASH__|$CHROMEAI_RUNTIME_HASH|g" \ -e "s|__OLLAMA_BASE_URL__|$OLLAMA_BASE_URL|g" \ -e "s|__OLLAMA_MODEL__|$OLLAMA_MODEL|g" \ -e "s|__OLLAMA_FAST_MODEL__|$OLLAMA_FAST_MODEL|g" \ diff --git a/server/nano_server.py b/server/nano_server.py index 13d35aa..fe3c40d 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -497,6 +497,43 @@ def require_binary_integrity(path: Path, expected_hash: str | None = None) -> No raise RuntimeError("runtime_integrity_check_failed") +RUNTIME_INTEGRITY_FILES = ( + "server/nano_server.py", + "server/ollama_client.py", + "server/extraction_presets.py", + "server/runtime_profiles.py", + "server/installation_manager.py", + "presets/extraction-presets.json", + "runtime/model_policy.json", + "runtime/model_registry.runtime.json", + "runtime/promotion_audit.json", +) + + +def runtime_integrity_digest(root: Path) -> str: + records: list[str] = [] + for relative_path in RUNTIME_INTEGRITY_FILES: + candidate = root / relative_path + if not candidate.is_file(): + raise RuntimeError(f"runtime integrity file missing: {relative_path}") + digest = hashlib.sha256(candidate.read_bytes()).hexdigest() + records.append(f"{relative_path}\t{digest}\n") + return hashlib.sha256("".join(records).encode("utf-8")).hexdigest() + + +def require_runtime_integrity(root: Path, expected_hash: str | None = None) -> None: + try: + digest = runtime_integrity_digest(root) + except (OSError, RuntimeError) as exc: + print(str(exc)) + raise RuntimeError("runtime_integrity_check_failed") from exc + if expected_hash and ( + re.fullmatch(r"[0-9a-f]{64}", expected_hash) is None or digest != expected_hash + ): + print(f"runtime hash mismatch: {digest} != {expected_hash}") + raise RuntimeError("runtime_integrity_check_failed") + + def ensure_dirs(path: Path): path.mkdir(parents=True, exist_ok=True) @@ -2086,6 +2123,7 @@ def main(): binary_path = Path(args.binary_path) if args.binary_path else Path(__file__) expected = args.binary_hash or os.environ.get('CHROMEAI_BINARY_HASH') require_binary_integrity(binary_path, expected) + require_runtime_integrity(PROJECT_ROOT, os.environ.get('CHROMEAI_RUNTIME_HASH')) auto_pull = os.environ.get('CHROMEAI_AUTO_PULL_MODELS', '0').strip().lower() not in {'0', 'false', 'no'} if auto_pull: diff --git a/server/ollama_client.py b/server/ollama_client.py index de7578b..3b710f3 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -324,6 +324,10 @@ def _request_json(self, path: str, payload: dict[str, Any] | None = None) -> Any raise OllamaError(f"{path} returned {e.code}: {raw or e.reason}") from e except URLError as e: raise OllamaError(f"cannot reach Ollama at {self.config.base_url}: {e.reason}") from e + except TimeoutError as e: + raise OllamaError( + f"Ollama request timed out after {self.config.timeout_seconds:g}s at {path}" + ) from e try: return json.loads(raw) if raw else {} diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index fe0dcc7..c333f10 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -64,15 +64,18 @@ test('macOS service installs outside the source checkout before launchd executio assert.match(installer, /ollama_client\.py/); assert.match(installer, /extraction_presets\.py/); assert.match(installer, /runtime_profiles\.py/); + assert.match(installer, /installation_manager\.py/); assert.match(installer, /presets\/extraction-presets\.json/); assert.match(installer, /model_policy\.json/); assert.match(installer, /model_registry\.runtime\.json/); + assert.match(installer, /CHROMEAI_RUNTIME_HASH/); assert.match(installer, /sys\.version_info >= \(3, 9\)/); assert.match(launchAgent, /__PYTHON_BIN__/); assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); assert.match(installer, /install -m 0555 "\$ROOT\/server\/\$module" "\$INSTALL_DIR\/\$module"/); assert.match(launchAgent, /__INSTALLED_BINARY__/); assert.match(launchAgent, /__INSTALL_DIR__/); + assert.match(launchAgent, /CHROMEAI_RUNTIME_HASH/); assert.doesNotMatch(launchAgent, /__PROJECT_ROOT__/); }); @@ -86,6 +89,7 @@ test('production macOS package also binds launchd to Python 3.9 or newer', () => assert.match(launchAgent, /__PYTHON_BIN__/); assert.match(launchAgent, /--state-dir/); assert.match(launchAgent, /CHROMEAI_RUNTIME_STATE_DIR/); + assert.match(launchAgent, /CHROMEAI_RUNTIME_HASH/); assert.match(launchAgent, /CHROMEAI_MAX_CONCURRENT_OPERATIONS<\/key>1/); assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); }); diff --git a/tests/server/test_bridge_origin.py b/tests/server/test_bridge_origin.py index 918be33..cb9d74d 100644 --- a/tests/server/test_bridge_origin.py +++ b/tests/server/test_bridge_origin.py @@ -9,6 +9,7 @@ import unittest from http.server import ThreadingHTTPServer from pathlib import Path +from unittest.mock import patch ROOT = Path(__file__).resolve().parents[2] SERVER_DIR = ROOT / "server" @@ -31,9 +32,16 @@ def tearDownClass(cls) -> None: cls.server.server_close() cls.thread.join(timeout=2) - def request(self, method: str, path: str, *, headers: dict[str, str] | None = None): + def request( + self, + method: str, + path: str, + *, + body: bytes | None = None, + headers: dict[str, str] | None = None, + ): connection = http.client.HTTPConnection("127.0.0.1", self.server.server_port, timeout=2) - connection.request(method, path, headers=headers or {}) + connection.request(method, path, body=body, headers=headers or {}) response = connection.getresponse() body = response.read() response_headers = dict(response.getheaders()) @@ -61,6 +69,19 @@ def test_rejects_oversized_request_before_reading_body(self) -> None: self.assertEqual(status, 413) self.assertEqual(json.loads(body)["error"]["code"], "request_too_large") + def test_ollama_transport_timeout_returns_explicit_service_error(self) -> None: + payload = json.dumps({"text": "selected text"}).encode("utf-8") + with patch("ollama_client.urlopen", side_effect=TimeoutError("timed out")): + status, _headers, body = self.request( + "POST", + "/extract", + body=payload, + headers={"Content-Type": "application/json"}, + ) + + self.assertEqual(status, 503) + self.assertEqual(json.loads(body)["error"]["code"], "ollama_unavailable") + if __name__ == "__main__": unittest.main() diff --git a/tests/server/test_capability_truth.py b/tests/server/test_capability_truth.py index a4edc79..8e2e89e 100644 --- a/tests/server/test_capability_truth.py +++ b/tests/server/test_capability_truth.py @@ -35,6 +35,40 @@ def test_binary_integrity_mismatch_stops_startup_contract(self): with self.assertRaisesRegex(RuntimeError, "runtime_integrity_check_failed"): nano_server.require_binary_integrity(binary, expected) + def test_runtime_integrity_covers_imported_code_and_policy(self): + relative_paths = ( + "server/nano_server.py", + "server/ollama_client.py", + "server/extraction_presets.py", + "server/runtime_profiles.py", + "server/installation_manager.py", + "presets/extraction-presets.json", + "runtime/model_policy.json", + "runtime/model_registry.runtime.json", + "runtime/promotion_audit.json", + ) + self.assertTrue( + hasattr(nano_server, "require_runtime_integrity"), + "runtime tree has no integrity verifier", + ) + + with tempfile.TemporaryDirectory() as temp_dir: + root = Path(temp_dir) + records = [] + for relative_path in relative_paths: + target = root / relative_path + target.parent.mkdir(parents=True, exist_ok=True) + target.write_text(f"trusted:{relative_path}\n", encoding="utf-8") + records.append(f"{relative_path}\t{sha256(target.read_bytes()).hexdigest()}\n") + + expected = sha256("".join(records).encode("utf-8")).hexdigest() + nano_server.require_runtime_integrity(root, expected) + + (root / "server/ollama_client.py").write_text("tampered\n", encoding="utf-8") + with redirect_stdout(io.StringIO()): + with self.assertRaisesRegex(RuntimeError, "runtime_integrity_check_failed"): + nano_server.require_runtime_integrity(root, expected) + def test_runtime_feedback_uses_explicit_mutable_state_directory(self): with tempfile.TemporaryDirectory() as temp_dir: state_dir = Path(temp_dir) / "state" diff --git a/tests/server/test_ollama_client.py b/tests/server/test_ollama_client.py index c622c2a..6c6e1c3 100644 --- a/tests/server/test_ollama_client.py +++ b/tests/server/test_ollama_client.py @@ -5,6 +5,7 @@ import sys import unittest from pathlib import Path +from unittest.mock import patch ROOT = Path(__file__).resolve().parents[2] SERVER_DIR = ROOT / "server" @@ -43,6 +44,26 @@ def _request_json(self, path: str, payload=None): class OllamaClientContractTests(unittest.TestCase): + def test_transport_timeout_is_normalized_as_ollama_error(self) -> None: + client = OllamaClient(OllamaConfig( + base_url="http://127.0.0.1:11434", + model="gemma4:e2b-it-qat", + embed_model="nomic-embed-text-v2-moe:latest", + timeout_seconds=45, + num_ctx=16384, + seed=42, + )) + + with patch("ollama_client.urlopen", side_effect=TimeoutError("timed out")): + try: + client._request_json("/api/generate", {"model": client.config.model}) + except OllamaError as exc: + self.assertRegex(str(exc), r"timed out after 45s.*api/generate") + except TimeoutError: + self.fail("transport TimeoutError escaped the Ollama client boundary") + else: + self.fail("expected an OllamaError") + def test_explicit_model_is_not_silently_substituted(self) -> None: client = StubOllamaClient({}) self.assertEqual( From 38390b760cc5844df1287d0df847e8602660e45b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 00:44:23 +0200 Subject: [PATCH 60/66] fix(deps): pin patched humanfs node runtime Force the ESLint filesystem helper to 0.16.8 so build tooling cannot follow attacker-controlled symlinks outside a copied source tree. The lockfile is regenerated and verified on Hetzner in the following commit. --- pnpm-lock.yaml | 26 ++++++++++++++++++-------- pnpm-workspace.yaml | 1 + 2 files changed, 19 insertions(+), 8 deletions(-) diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml index 093c74b..40fd365 100644 --- a/pnpm-lock.yaml +++ b/pnpm-lock.yaml @@ -5,6 +5,7 @@ settings: excludeLinksFromLockfile: false overrides: + '@humanfs/node': 0.16.8 brace-expansion: 5.0.9 importers: @@ -233,12 +234,16 @@ packages: resolution: {integrity: sha512-iH1B076HoAshH1mLpHMgwdGeTs0CYwL0SPMkGuSebZrwBp16v415e9NZXg2jtrqPVQjf6IANe2Vtlr5KswtcZQ==} engines: {node: ^20.19.0 || ^22.13.0 || >=24} - '@humanfs/core@0.19.1': - resolution: {integrity: sha512-5DyQ4+1JEUzejeK1JGICcideyfUbGixgS9jNgex5nqkW+cY7WZhxBigmieN5Qnw9ZosSNVC9KQKyb+GUaGyKUA==} + '@humanfs/core@0.19.2': + resolution: {integrity: sha512-UhXNm+CFMWcbChXywFwkmhqjs3PRCmcSa/hfBgLIb7oQ5HNb1wS0icWsGtSAUNgefHeI+eBrA8I1fxmbHsGdvA==} engines: {node: '>=18.18.0'} - '@humanfs/node@0.16.7': - resolution: {integrity: sha512-/zUx+yOsIrG4Y43Eh2peDeKCxlRt/gET6aHfaKpuq267qXdYDFViVHfMaLyygZOnl0kGWxFIgsBy8QFuTLUXEQ==} + '@humanfs/node@0.16.8': + resolution: {integrity: sha512-gE1eQNZ3R++kTzFUpdGlpmy8kDZD/MLyHqDwqjkVQI0JMdI1D51sy1H958PNXYkM2rAac7e5/CnIKZrHtPh3BQ==} + engines: {node: '>=18.18.0'} + + '@humanfs/types@0.15.0': + resolution: {integrity: sha512-ZZ1w0aoQkwuUuC7Yf+7sdeaNfqQiiLcSRbfI08oAxqLtpXQr9AIVX7Ay7HLDuiLYAaFPu8oBYNq/QIi9URHJ3Q==} engines: {node: '>=18.18.0'} '@humanwhocodes/module-importer@1.0.1': @@ -732,13 +737,18 @@ snapshots: '@eslint/core': 1.1.1 levn: 0.4.1 - '@humanfs/core@0.19.1': {} + '@humanfs/core@0.19.2': + dependencies: + '@humanfs/types': 0.15.0 - '@humanfs/node@0.16.7': + '@humanfs/node@0.16.8': dependencies: - '@humanfs/core': 0.19.1 + '@humanfs/core': 0.19.2 + '@humanfs/types': 0.15.0 '@humanwhocodes/retry': 0.4.3 + '@humanfs/types@0.15.0': {} + '@humanwhocodes/module-importer@1.0.1': {} '@humanwhocodes/retry@0.4.3': {} @@ -938,7 +948,7 @@ snapshots: '@eslint/config-helpers': 0.5.3 '@eslint/core': 1.1.1 '@eslint/plugin-kit': 0.6.1 - '@humanfs/node': 0.16.7 + '@humanfs/node': 0.16.8 '@humanwhocodes/module-importer': 1.0.1 '@humanwhocodes/retry': 0.4.3 '@types/estree': 1.0.8 diff --git a/pnpm-workspace.yaml b/pnpm-workspace.yaml index f778b49..543346a 100644 --- a/pnpm-workspace.yaml +++ b/pnpm-workspace.yaml @@ -2,6 +2,7 @@ packages: - '.' overrides: + '@humanfs/node': 0.16.8 brace-expansion: 5.0.9 onlyBuiltDependencies: From 2660b892c3f21ea4066cb9e0b53471db3d0bdd4b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 00:57:06 +0200 Subject: [PATCH 61/66] ci(security): move CodeQL workflow to v4 Update init, autobuild, and analyze together to remove the December 2026 v3 deprecation and the legacy Node 20 action runtime warning. --- .github/workflows/codeql.yml | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/.github/workflows/codeql.yml b/.github/workflows/codeql.yml index d2d3785..a2a49ee 100644 --- a/.github/workflows/codeql.yml +++ b/.github/workflows/codeql.yml @@ -30,7 +30,7 @@ jobs: - uses: actions/checkout@v7 - name: Initialize CodeQL - uses: github/codeql-action/init@v3 + uses: github/codeql-action/init@v4 with: languages: ${{ matrix.language }} @@ -59,7 +59,7 @@ jobs: run: pnpm install --frozen-lockfile - name: Autobuild - uses: github/codeql-action/autobuild@v3 + uses: github/codeql-action/autobuild@v4 - name: Perform CodeQL Analysis - uses: github/codeql-action/analyze@v3 + uses: github/codeql-action/analyze@v4 From 05dbc477f3ea647c5768841f1019e21a4f86448d Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 01:10:22 +0200 Subject: [PATCH 62/66] docs(governance): reconcile final candidate evidence Record the admitted implementation parent, physical diagnostic boundary, protected Store environment state, and remaining external authorities without promoting the project to production completion. --- .../20-checkpoint.md | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md index c30b207..ebafdef 100644 --- a/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md +++ b/docs/aegis/work/2026-08-30-production-abuse-gate/20-checkpoint.md @@ -2,13 +2,13 @@ ## TodoCheckpointDraft -- Current todo: qualify the physical-smoke follow-up on Hetzner and GitHub, preserve the independent-review gate, and retain every unavailable production authority without laundering it into code completion. -- Completed: Linear global runtime laws, SelectPilot project truth, and GitHub PR state were re-read before mutation; the full-system gap/blind-spot/negative-space audit was refreshed; the signed intent remains valid. PR #22 source `93334b73e58bc122947eb6b8b866f3a2e65e0e50` passed all 9 exact-head GitHub checks in CI run `33812111260` and CodeQL run `33812111254`. +- Current todo: preserve the independent-review gate and retain every unavailable production authority without laundering the admitted repository candidate into production completion. +- Completed: Linear global runtime laws, SelectPilot project truth, and GitHub PR state were re-read before mutation; the full-system gap/blind-spot/negative-space audit was refreshed; the signed intent remains valid. The implementation/workflow parent `2660b892c3f21ea4066cb9e0b53471db3d0bdd4b` passed all 9 exact-head GitHub checks in CI run `33815452340` and CodeQL v4 run `33815452342` before this documentation-only reconciliation. - Repository repairs: Store packaging now requires an explicit external entitlement public identity and exact clean SHA; ordinary CI proves the missing-identity failure path; mutable helper state is user-owned; local inference concurrency is bounded; exact model/context routing survives the agent pipeline and is visible in the extension. Physical inspection also found and repaired two installer/runtime gaps: transport `TimeoutError` now becomes explicit bridge HTTP 503, and both local and packaged LaunchAgents bind an aggregate hash over all active runtime code and policy. The local installer now includes the previously omitted `installation_manager.py` dependency. -- Remote evidence: Hetzner and exact-head GitHub passed lint, typecheck/build parity, 50 Node tests, 73 server tests, 7 Playwright journeys, privacy, billing, manifest/assets, negative Store packaging, frontier traceability, runtime-policy validation, stress contracts, and exact-SHA code admission. The admission report correctly remained `runtime_verified=false`, `physical_hardware_verified=false`, and `full_system_ok=false`. +- Remote evidence: Hetzner and exact-head GitHub passed lint, typecheck/build parity, 50 Node tests, 76 server tests, 7 Playwright journeys, privacy, billing, manifest/assets, negative Store packaging, frontier traceability, runtime-policy validation, stress contracts, and exact-SHA code admission. Dependency Review reported no vulnerable package at severity low or higher after pinning `brace-expansion@5.0.9` and `@humanfs/node@0.16.8`; both CodeQL v4 languages passed with no open code-scanning alert. Admission artifact `9916562422` has SHA-256 `fc27b9374945e1df28580aa91f567491941963b7d1a6a505d03b1c58f23f4525` and correctly remains `runtime_verified=false`, `physical_hardware_verified=false`, and `full_system_ok=false`. - Physical M4 evidence: the explicit local prewarm completed in `130.599s` (`129.834s` model load) from external model storage. Once warm, two identical `action_brief` extractions completed in `3.722s` and `3.134s`, selected `gemma4:e2b-it-qat` at 16,384 context with `smallest_qualified_structured_model`, and produced identical canonical digest `8283d9866499bef776bc4d171b94d772944a04330ffed329774baa438c55b6d4`. An explicitly missing model returned degraded health and HTTP 503 without substitution; restoring the exact model recovered extraction in `3.446s`. -- External truth: PR #22 still requires one eligible independent approval; production DNS/TLS, Paddle credentials/lifecycle, authorized entitlement signer, Developer ID Installer/notarization, exact Store upload/publication, full browser first-use/export evidence, and remaining declared Apple hardware profiles remain outside repository authority. -- Next: obtain full remote regression and a new signed exact-head GitHub result for the physical-smoke repair, then execute protected merge and production gates only when their real authorities exist. +- External truth: PR #22 still requires one eligible independent approval. GitHub environment `chrome-web-store-production` now accepts protected branches only and forbids admin bypass, but holds zero secrets. Production DNS/TLS, Paddle credentials/lifecycle, authorized entitlement signer, Developer ID Installer/notarization, exact Store upload/publication, full browser first-use/export evidence, exact-final-SHA physical acceptance, and remaining declared Apple hardware profiles remain outside repository authority. +- Next: obtain an eligible independent approval, then execute protected merge and the production/physical gates only when their real authorities exist. ## ResumeStateHint @@ -19,5 +19,5 @@ Resume from this checkpoint, re-read `10-intent.md`, verify `intent.json`, compa - Intent alignment: yes. - Scope alignment: yes. - New fallback or duplicate owner: no. -- Evidence sufficiency: repository hardening and one physical M4 structured-runtime slice are qualified; final follow-up GitHub admission, complete physical/browser runtime, protected review, and production activation evidence are still required. -- Decision: continue through remote and GitHub admission; fail closed on every full-system completion claim until its authority is present. +- Evidence sufficiency: exact-head repository hardening and one earlier runtime-identical physical M4 structured-runtime slice are qualified; exact-final-SHA physical/browser runtime, protected review, and production activation evidence are still required. +- Decision: repository candidate admitted; fail closed on every merge, physical acceptance, production, and full-system completion claim until its authority is present. From 14b862b667bb214266a845cd0f694c98449f5b24 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 06:49:35 +0200 Subject: [PATCH 63/66] fix(extension): enforce local storage and tab-bound opening Keep extension preferences device-local, bind the side panel to the active tab before opening, surface restricted-page failures, and add verified plain-text export. Verified: Hetzner pnpm lint && pnpm test && pnpm test:e2e --- panel/panel.js | 9 ++- panel/panel.ts | 10 ++- popup/popup.css | 4 + popup/popup.html | 1 + popup/popup.js | 15 +++- popup/popup.ts | 15 +++- tests/e2e/panel-harness.html | 4 - tests/e2e/panel-user-flow.spec.mjs | 7 ++ tests/panel/local-settings-boundary.test.mjs | 10 ++- tests/panel/popup-open-contract.test.mjs | 79 ++++++++++++++++++++ 10 files changed, 141 insertions(+), 13 deletions(-) create mode 100644 tests/panel/popup-open-contract.test.mjs diff --git a/panel/panel.js b/panel/panel.js index c65c7bb..ba73bf9 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -682,6 +682,13 @@ function renderExports({ markdown, json, basename = 'selectpilot' }) { setStatus('Markdown downloaded'); }); actions.append(downloadMarkdown); + const downloadPlainText = document.createElement('button'); + downloadPlainText.textContent = 'Download .txt'; + downloadPlainText.addEventListener('click', () => { + triggerDownload(markdown, `${basename}.txt`, 'text/plain;charset=utf-8'); + setStatus('Plain text downloaded'); + }); + actions.append(downloadPlainText); } if (json && Object.keys(json).length > 0) { const jsonText = JSON.stringify(json, null, 2); @@ -1431,7 +1438,7 @@ function bindActions() { populatePresetOptions(); bindActions(); async function initialize() { - const storedSettings = await chrome.storage.sync.get('selectpilot_settings'); + const storedSettings = await chrome.storage.local.get('selectpilot_settings'); const settings = (storedSettings.selectpilot_settings || {}); const textSize = typeof settings.textSize === 'string' ? settings.textSize : 'standard'; document.body.dataset.controlSide = settings.controlSide === 'left' ? 'left' : 'right'; diff --git a/panel/panel.ts b/panel/panel.ts index 969263f..d58d3c5 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -836,6 +836,14 @@ function renderExports({ setStatus('Markdown downloaded'); }); actions.append(downloadMarkdown); + + const downloadPlainText = document.createElement('button'); + downloadPlainText.textContent = 'Download .txt'; + downloadPlainText.addEventListener('click', () => { + triggerDownload(markdown, `${basename}.txt`, 'text/plain;charset=utf-8'); + setStatus('Plain text downloaded'); + }); + actions.append(downloadPlainText); } if (json && Object.keys(json).length > 0) { @@ -1591,7 +1599,7 @@ populatePresetOptions(); bindActions(); async function initialize() { - const storedSettings = await chrome.storage.sync.get('selectpilot_settings'); + const storedSettings = await chrome.storage.local.get('selectpilot_settings'); const settings = (storedSettings.selectpilot_settings || {}) as Record; const textSize = typeof settings.textSize === 'string' ? settings.textSize : 'standard'; document.body.dataset.controlSide = settings.controlSide === 'left' ? 'left' : 'right'; diff --git a/popup/popup.css b/popup/popup.css index a58eea6..db81aad 100644 --- a/popup/popup.css +++ b/popup/popup.css @@ -46,3 +46,7 @@ button { box-shadow: 0 14px 30px rgba(56, 189, 248, 0.2); } button:active { opacity: 0.9; } + +.popup-status { + color: #fca5a5; +} diff --git a/popup/popup.html b/popup/popup.html index efb8b5f..2825de5 100644 --- a/popup/popup.html +++ b/popup/popup.html @@ -11,6 +11,7 @@

    SelectPilot

    Open the side panel to turn selected text into summaries, rewrites, or next steps with Ollama running locally.

    + diff --git a/popup/popup.js b/popup/popup.js index 3ea4371..8dfde5f 100644 --- a/popup/popup.js +++ b/popup/popup.js @@ -2,23 +2,32 @@ // spec_ref: "frontend_state_contract" import { log } from '../utils/logger.js'; const btn = document.getElementById('open-panel'); +const statusEl = document.getElementById('popup-status'); async function injectIntoActiveTab() { const [tab] = await chrome.tabs.query({ active: true, currentWindow: true }); - if (!tab?.id) + if (!tab?.id || tab.windowId === undefined) throw new Error('No active tab is available'); await chrome.scripting.executeScript({ target: { tabId: tab.id }, files: ['content/content-script.bundle.js'], }); + await chrome.sidePanel.setOptions({ tabId: tab.id, path: 'panel/panel.html', enabled: true }); + await chrome.sidePanel.open({ windowId: tab.windowId }); } btn?.addEventListener('click', async () => { + if (statusEl) { + statusEl.textContent = ''; + statusEl.hidden = true; + } try { await injectIntoActiveTab(); - await chrome.sidePanel.open({ windowId: chrome.windows.WINDOW_ID_CURRENT }); - await chrome.sidePanel.setOptions({ path: 'panel/panel.html', enabled: true }); window.close(); } catch (e) { log('popup', 'failed to open side panel', e); + if (statusEl) { + statusEl.textContent = 'SelectPilot cannot run on this page. Open a regular webpage and try again.'; + statusEl.hidden = false; + } } }); diff --git a/popup/popup.ts b/popup/popup.ts index aec88d1..f2b8339 100644 --- a/popup/popup.ts +++ b/popup/popup.ts @@ -3,23 +3,32 @@ import { log } from '../utils/logger.js'; const btn = document.getElementById('open-panel'); +const statusEl = document.getElementById('popup-status'); async function injectIntoActiveTab(): Promise { const [tab] = await chrome.tabs.query({ active: true, currentWindow: true }); - if (!tab?.id) throw new Error('No active tab is available'); + if (!tab?.id || tab.windowId === undefined) throw new Error('No active tab is available'); await chrome.scripting.executeScript({ target: { tabId: tab.id }, files: ['content/content-script.bundle.js'], }); + await chrome.sidePanel.setOptions({ tabId: tab.id, path: 'panel/panel.html', enabled: true }); + await chrome.sidePanel.open({ windowId: tab.windowId }); } btn?.addEventListener('click', async () => { + if (statusEl) { + statusEl.textContent = ''; + statusEl.hidden = true; + } try { await injectIntoActiveTab(); - await chrome.sidePanel.open({ windowId: chrome.windows.WINDOW_ID_CURRENT }); - await chrome.sidePanel.setOptions({ path: 'panel/panel.html', enabled: true }); window.close(); } catch (e) { log('popup', 'failed to open side panel', e); + if (statusEl) { + statusEl.textContent = 'SelectPilot cannot run on this page. Open a regular webpage and try again.'; + statusEl.hidden = false; + } } }); diff --git a/tests/e2e/panel-harness.html b/tests/e2e/panel-harness.html index 2c53eae..f84d5cd 100644 --- a/tests/e2e/panel-harness.html +++ b/tests/e2e/panel-harness.html @@ -254,10 +254,6 @@

    Ready

    async create() {}, }, storage: { - sync: { - async get() { return {}; }, - async set() {}, - }, local: { _store: {}, async get(key) { diff --git a/tests/e2e/panel-user-flow.spec.mjs b/tests/e2e/panel-user-flow.spec.mjs index 253cf8a..03ef5f9 100644 --- a/tests/e2e/panel-user-flow.spec.mjs +++ b/tests/e2e/panel-user-flow.spec.mjs @@ -1,6 +1,7 @@ // module_name: tests_e2e_panel-user-flow_spec_mjs // spec_ref: "testing_strategy.integration_tests" import { test, expect } from '@playwright/test'; +import { readFile } from 'node:fs/promises'; import { resolve } from 'node:path'; import { pathToFileURL } from 'node:url'; @@ -45,6 +46,12 @@ test('unlocked first-run uses the dedicated example route and resolves once', as await expect(page.locator('#workflow')).toContainText('Publishing waits for both checks'); await expect(page.locator('#exports')).toContainText('Copy Markdown'); await expect(page.locator('#exports')).toContainText('Copy JSON'); + await expect(page.getByRole('button', { name: 'Download .txt' })).toBeVisible(); + const downloadPromise = page.waitForEvent('download'); + await page.getByRole('button', { name: 'Download .txt' }).click(); + const download = await downloadPromise; + expect(download.suggestedFilename()).toBe('selectpilot-first-result.txt'); + expect(await readFile(await download.path(), 'utf8')).toContain('Maya: privacy check by Thursday'); await expect(page.locator('#btn-first-run-example')).toHaveCount(0); const demoMessages = await page.evaluate(() => globalThis.__messageLog.filter((msg) => msg.type === 'panel:extract_demo')); diff --git a/tests/panel/local-settings-boundary.test.mjs b/tests/panel/local-settings-boundary.test.mjs index 3aefbd8..f57fbfd 100644 --- a/tests/panel/local-settings-boundary.test.mjs +++ b/tests/panel/local-settings-boundary.test.mjs @@ -2,11 +2,19 @@ // spec_ref: "privacy_and_debug_policy" import assert from 'node:assert/strict'; import { readFile } from 'node:fs/promises'; +import { fileURLToPath } from 'node:url'; import test from 'node:test'; +import { collectRuntimeFiles } from '../../scripts/package-chrome-store.mjs'; test('extension preferences remain device-local and never use Chrome sync', async () => { const source = await readFile(new URL('../../options/options.ts', import.meta.url), 'utf8'); assert.match(source, /chrome\.storage\.local\.get/); assert.match(source, /chrome\.storage\.local\.set/); - assert.doesNotMatch(source, /chrome\.storage\.sync/); + + const projectRoot = fileURLToPath(new URL('../../', import.meta.url)); + const runtimeFiles = (await collectRuntimeFiles(projectRoot)).filter((file) => /\.(?:html|js)$/.test(file)); + for (const runtimeFile of runtimeFiles) { + const runtimeSource = await readFile(new URL(`../../${runtimeFile}`, import.meta.url), 'utf8'); + assert.doesNotMatch(runtimeSource, /chrome\.storage\.sync/, `${runtimeFile} must not use Chrome Sync`); + } }); diff --git a/tests/panel/popup-open-contract.test.mjs b/tests/panel/popup-open-contract.test.mjs new file mode 100644 index 0000000..3abe195 --- /dev/null +++ b/tests/panel/popup-open-contract.test.mjs @@ -0,0 +1,79 @@ +// module_name: popup_open_contract_tests +// spec_ref: "frontend_state_contract" +import assert from 'node:assert/strict'; +import test from 'node:test'; + +test('toolbar popup binds the side panel to the active tab before opening it', async () => { + const calls = []; + let clickHandler; + const status = { textContent: '', hidden: true }; + const button = { + addEventListener: (type, handler) => { + if (type === 'click') clickHandler = handler; + }, + }; + globalThis.document = { + getElementById: (id) => id === 'open-panel' ? button : id === 'popup-status' ? status : null, + }; + globalThis.chrome = { + tabs: { query: async () => [{ id: 42, windowId: 7 }] }, + scripting: { executeScript: async ({ target }) => calls.push(`inject:${target.tabId}`) }, + sidePanel: { + setOptions: async ({ tabId, path, enabled }) => calls.push(`set:${tabId}:${path}:${enabled}`), + open: async ({ windowId }) => calls.push(`open:${windowId}`), + }, + windows: { WINDOW_ID_CURRENT: -2 }, + }; + globalThis.window = { close: () => calls.push('close') }; + + try { + await import(`../../popup/popup.js?test=${Date.now()}`); + assert.equal(typeof clickHandler, 'function'); + await clickHandler(); + assert.deepEqual(calls, [ + 'inject:42', + 'set:42:panel/panel.html:true', + 'open:7', + 'close', + ]); + } finally { + delete globalThis.document; + delete globalThis.chrome; + delete globalThis.window; + } +}); + +test('toolbar popup keeps an actionable error visible when page injection fails', async () => { + let clickHandler; + let closed = false; + const status = { textContent: '', hidden: true }; + const button = { + addEventListener: (type, handler) => { + if (type === 'click') clickHandler = handler; + }, + }; + globalThis.document = { + getElementById: (id) => id === 'open-panel' ? button : id === 'popup-status' ? status : null, + }; + globalThis.chrome = { + tabs: { query: async () => [{ id: 42, windowId: 7 }] }, + scripting: { executeScript: async () => { throw new Error('restricted page'); } }, + sidePanel: { setOptions: async () => {}, open: async () => {} }, + }; + globalThis.window = { close: () => { closed = true; } }; + const originalConsoleError = console.error; + console.error = () => {}; + + try { + await import(`../../popup/popup.js?failure-test=${Date.now()}`); + await clickHandler(); + assert.equal(status.hidden, false); + assert.equal(status.textContent, 'SelectPilot cannot run on this page. Open a regular webpage and try again.'); + assert.equal(closed, false); + } finally { + console.error = originalConsoleError; + delete globalThis.document; + delete globalThis.chrome; + delete globalThis.window; + } +}); From 41afa0540bdbad89e577614e7940d45a811b6cf8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 07:08:07 +0200 Subject: [PATCH 64/66] perf(runtime): default to smallest qualified model Select the Fast E2B profile for automatic setup on every supported machine while keeping heavier profiles behind explicit operator choice. Verified: Hetzner pnpm lint && pnpm typecheck && pnpm test && pnpm test:e2e --- README.md | 2 +- docs/AI_RUNTIME_LANE_MATRIX.md | 1 + server/runtime_profiles.py | 24 +++++------------------- tests/panel/bootstrap-contract.test.mjs | 4 ++-- tests/server/test_runtime_profiles.py | 9 +++++++-- 5 files changed, 16 insertions(+), 24 deletions(-) diff --git a/README.md b/README.md index 999ad43..108092a 100644 --- a/README.md +++ b/README.md @@ -47,7 +47,7 @@ SelectPilot is built to keep the core workflow local-first and inspectable. pnpm setup:local ``` -This installs dependencies, builds the extension, installs/starts Ollama when needed, selects a hardware-safe local profile, prepares the exact local model bundle, installs the LaunchAgent, and refuses to finish unless the bridge and configured models are healthy. Extraction and summaries use the smallest qualified local model; broader agent tasks use the profile's general model. Structured generation uses a fixed zero-temperature sampling contract and seed so repeated requests follow the same decision path. +This installs dependencies, builds the extension, installs/starts Ollama when needed, selects the smallest qualified local profile by default, prepares the exact local model bundle, installs the LaunchAgent, and refuses to finish unless the bridge and configured models are healthy. Heavier profiles remain available only through explicit manual selection. Extraction and summaries use the smallest qualified local model; agent tasks use the selected profile's qualified general model. Structured generation uses a fixed zero-temperature sampling contract and seed so repeated requests follow the same decision path. 2. Open `chrome://extensions`, enable **Developer mode**, choose **Load unpacked**, and select this repository folder. 3. Highlight text on a page → open SelectPilot → click **Extract JSON**. diff --git a/docs/AI_RUNTIME_LANE_MATRIX.md b/docs/AI_RUNTIME_LANE_MATRIX.md index 25a66e6..fb97761 100644 --- a/docs/AI_RUNTIME_LANE_MATRIX.md +++ b/docs/AI_RUNTIME_LANE_MATRIX.md @@ -36,6 +36,7 @@ This contract defines which local runtime may perform each class of work without - No silent provider or model failover. - No remote inference fallback. +- Automatic setup selects the smallest qualified local profile; heavier profiles require explicit operator selection. - Gemini Nano cannot create, repair, or persist Golden JSON. - WebGPU is acceleration, never semantic authority. - Capability detection sends no selected text and creates no model session. diff --git a/server/runtime_profiles.py b/server/runtime_profiles.py index 5cd0b29..9878f97 100644 --- a/server/runtime_profiles.py +++ b/server/runtime_profiles.py @@ -102,25 +102,11 @@ def detect_system_snapshot() -> dict[str, Any]: def recommend_runtime_profile(system_snapshot: dict[str, Any] | None = None) -> dict[str, Any]: snapshot = system_snapshot or detect_system_snapshot() - memory_gb = snapshot.get("memory_gb") - machine = str(snapshot.get("machine") or "") - - if memory_gb is None: - profile = RUNTIME_PROFILES["fast"] - reason = "Memory could not be detected, so the smallest viable profile is safest." - elif memory_gb < 16: - profile = RUNTIME_PROFILES["fast"] - reason = "This machine benefits from the smallest viable profile for low-latency extraction." - elif memory_gb < 32: - profile = RUNTIME_PROFILES["balanced"] - reason = "This machine can comfortably handle the balanced profile without overprovisioning." - else: - profile = RUNTIME_PROFILES["balanced"] - reason = "Even on larger machines, balanced is the default because SelectPilot prioritizes fit-for-task over maximum model size." - - if machine.startswith("x86") and profile.key != "fast": - profile = RUNTIME_PROFILES["fast"] - reason = "Intel machines default to the fast profile unless you explicitly opt into heavier models." + profile = RUNTIME_PROFILES["fast"] + reason = ( + "SelectPilot automatically uses the smallest qualified local profile. " + "Heavier profiles require explicit operator selection." + ) return { "recommended_profile": profile.key, diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index c333f10..6fc8bfa 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -42,7 +42,7 @@ test('bootstrap plan exposes exact fail-closed Balanced runtime contract without assert.equal(plan.reason, 'Explicit profile selected by operator.'); }); -test('one-command setup preserves hardware-aware automatic profile selection', () => { +test('one-command setup preserves smallest-qualified automatic profile selection', () => { const packageJson = JSON.parse(readFileSync('./package.json', 'utf8')); assert.match(packageJson.scripts['setup:local'], /bootstrap-macos-local\.sh --profile auto$/); }); @@ -94,7 +94,7 @@ test('production macOS package also binds launchd to Python 3.9 or newer', () => assert.doesNotMatch(launchAgent, /\/usr\/bin\/python3/); }); -test('production macOS package preserves hardware-aware task routing', () => { +test('production macOS package preserves explicit task routing and profile truth', () => { const postinstall = readFileSync('./installer/macos/scripts/postinstall', 'utf8'); const launchAgent = readFileSync('./installer/macos/com.selectpilot.bridge.plist.template', 'utf8'); diff --git a/tests/server/test_runtime_profiles.py b/tests/server/test_runtime_profiles.py index 78835bf..9085f22 100644 --- a/tests/server/test_runtime_profiles.py +++ b/tests/server/test_runtime_profiles.py @@ -32,11 +32,16 @@ def test_unknown_profile_fails_explicitly(self) -> None: self.assertEqual(profile.generation_model, "gemma4:e2b-it-qat") self.assertEqual(profile.num_ctx, 16384) - def test_auto_recommendation_prefers_balanced_on_large_machines(self) -> None: + def test_auto_recommendation_uses_smallest_qualified_profile_on_large_machines(self) -> None: recommendation = recommend_runtime_profile( {"machine": "arm64", "memory_gb": 64, "platform": "darwin", "cpu_count": 10} ) - self.assertEqual(recommendation["recommended_profile"], "balanced") + self.assertEqual(recommendation["recommended_profile"], "fast") + self.assertIn("smallest qualified", recommendation["reason"].lower()) + auto_profile = get_runtime_profile(recommendation["recommended_profile"]) + self.assertEqual(required_generation_models(auto_profile), [ + ("gemma4:e2b-it-qat", 16384), + ]) def test_low_memory_and_intel_hardware_use_fast(self) -> None: low_memory = recommend_runtime_profile( From 8a26f54d5e554c90843d1b8571c50124b84d11c1 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 07:21:23 +0200 Subject: [PATCH 65/66] fix(panel): keep automatic profile on smallest model Verified: node --test tests/panel/runtime-profiles.test.mjs --- panel/panel.js | 4 ++-- panel/panel.ts | 14 ++++++++++-- panel/runtime-profiles.js | 4 ++++ panel/runtime-profiles.ts | 14 ++++++++++++ tests/panel/runtime-profiles.test.mjs | 32 +++++++++++++++++++++++++++ 5 files changed, 64 insertions(+), 4 deletions(-) create mode 100644 tests/panel/runtime-profiles.test.mjs diff --git a/panel/panel.js b/panel/panel.js index ba73bf9..a1120b4 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -7,7 +7,7 @@ import { getJSON, setJSON } from '../utils/storage.js'; import { endpoints } from '../api/endpoints.js'; import { getRuntimeMetaHealth, getRuntimeMetaStreamUrl } from '../api/nano-client.js'; import { EXTRACTION_PRESETS, getExtractionPreset } from './extraction-presets.js'; -import { getRuntimeProfile, RUNTIME_PROFILES } from './runtime-profiles.js'; +import { getRuntimeProfile, resolveRuntimeProfileKey, RUNTIME_PROFILES, } from './runtime-profiles.js'; import { buildKnowledgePackage } from './knowledge-connectors.js'; import { applyRuntimeEvent, setIntent, setSelectionContext, setVisiblePanels } from './state/runtimeStore.js'; import { loadBottleneckReport, loadDeterminismReport, loadFrontierReport } from './state/reportStore.js'; @@ -575,7 +575,7 @@ function getEntitlementCacheState(snapshot) { return 'stale'; } function getEffectiveRecommendedProfileKey() { - return benchmarkSnapshot?.recommended_profile || runtimeProfilesPayload.recommended_profile; + return resolveRuntimeProfileKey(privacyProofSnapshot?.runtime_profile, runtimeProfilesPayload.recommended_profile, benchmarkSnapshot); } function getEffectiveRecommendedProfile() { return getRuntimeProfile(getEffectiveRecommendedProfileKey()); diff --git a/panel/panel.ts b/panel/panel.ts index d58d3c5..b6b058a 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -7,7 +7,12 @@ import { getJSON, setJSON } from '../utils/storage.js'; import { endpoints } from '../api/endpoints.js'; import { getRuntimeMetaHealth, getRuntimeMetaStreamUrl, type RuntimeMetaEvent, type IntentCompileResponse } from '../api/nano-client.js'; import { EXTRACTION_PRESETS, getExtractionPreset, type ExtractionPresetKey } from './extraction-presets.js'; -import { getRuntimeProfile, RUNTIME_PROFILES, type RuntimeProfile } from './runtime-profiles.js'; +import { + getRuntimeProfile, + resolveRuntimeProfileKey, + RUNTIME_PROFILES, + type RuntimeProfile, +} from './runtime-profiles.js'; import { buildKnowledgePackage, type KnowledgeTarget, type MemoryLedgerEntry } from './knowledge-connectors.js'; import { applyRuntimeEvent, setIntent, setSelectionContext, setVisiblePanels } from './state/runtimeStore.js'; import { loadBottleneckReport, loadDeterminismReport, loadFrontierReport } from './state/reportStore.js'; @@ -103,6 +108,7 @@ type BenchmarkSnapshot = { type PrivacyProofSnapshot = { ok: boolean; privacy_mode: string; + runtime_profile?: string; generated_at?: string; outbound_observation?: { external_calls_registered?: boolean; @@ -699,7 +705,11 @@ function getEntitlementCacheState(snapshot: EntitlementSnapshot): 'fresh' | 'cac } function getEffectiveRecommendedProfileKey(): string { - return benchmarkSnapshot?.recommended_profile || runtimeProfilesPayload.recommended_profile; + return resolveRuntimeProfileKey( + privacyProofSnapshot?.runtime_profile, + runtimeProfilesPayload.recommended_profile, + benchmarkSnapshot, + ); } function getEffectiveRecommendedProfile() { diff --git a/panel/runtime-profiles.js b/panel/runtime-profiles.js index be37ae8..e03709b 100644 --- a/panel/runtime-profiles.js +++ b/panel/runtime-profiles.js @@ -47,3 +47,7 @@ export const RUNTIME_PROFILES = [ export function getRuntimeProfile(key) { return RUNTIME_PROFILES.find((profile) => profile.key === key) || RUNTIME_PROFILES[0]; } +// @spec_ref frontend_state_contract +export function resolveRuntimeProfileKey(configuredProfile, backendRecommendedProfile, benchmark) { + return getRuntimeProfile(configuredProfile || benchmark?.auto_profile || backendRecommendedProfile).key; +} diff --git a/panel/runtime-profiles.ts b/panel/runtime-profiles.ts index f072831..6bb05f3 100644 --- a/panel/runtime-profiles.ts +++ b/panel/runtime-profiles.ts @@ -18,6 +18,11 @@ export type RuntimeProfile = { is_default_auto?: boolean; }; +export type BenchmarkProfileHint = { + auto_profile?: string | null; + recommended_profile?: string | null; +}; + export const RUNTIME_PROFILES: RuntimeProfile[] = [ { key: 'fast', @@ -68,3 +73,12 @@ export const RUNTIME_PROFILES: RuntimeProfile[] = [ export function getRuntimeProfile(key: string | null | undefined): RuntimeProfile { return RUNTIME_PROFILES.find((profile) => profile.key === key) || RUNTIME_PROFILES[0]; } + +// @spec_ref frontend_state_contract +export function resolveRuntimeProfileKey( + configuredProfile: string | null | undefined, + backendRecommendedProfile: string | null | undefined, + benchmark: BenchmarkProfileHint | null | undefined, +): RuntimeProfileKey { + return getRuntimeProfile(configuredProfile || benchmark?.auto_profile || backendRecommendedProfile).key; +} diff --git a/tests/panel/runtime-profiles.test.mjs b/tests/panel/runtime-profiles.test.mjs new file mode 100644 index 0000000..8f26b41 --- /dev/null +++ b/tests/panel/runtime-profiles.test.mjs @@ -0,0 +1,32 @@ +// module_name: panel_runtime_profiles_test +// spec_ref: "frontend_state_contract" + +import assert from 'node:assert/strict'; +import test from 'node:test'; + +import { resolveRuntimeProfileKey } from '../../panel/runtime-profiles.js'; + +test('automatic profile truth ignores a larger benchmark suggestion', () => { + assert.equal(resolveRuntimeProfileKey(undefined, 'fast', { recommended_profile: 'advanced' }), 'fast'); +}); + +test('automatic profile truth accepts the benchmark auto profile contract', () => { + assert.equal(resolveRuntimeProfileKey(undefined, 'fast', { + auto_profile: 'fast', + recommended_profile: 'balanced', + }), 'fast'); +}); + +test('an explicit active profile remains visible without changing auto defaults', () => { + assert.equal(resolveRuntimeProfileKey('balanced', 'fast', { + auto_profile: 'fast', + recommended_profile: 'advanced', + }), 'balanced'); +}); + +test('automatic profile truth fails safely to the smallest known profile', () => { + assert.equal(resolveRuntimeProfileKey('invalid', 'unknown', { + auto_profile: 'invalid', + recommended_profile: 'advanced', + }), 'fast'); +}); From 0e5ec938cbf1e12b22bda6efb4866771aa4273d4 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Kristoffer=20S=C3=B6dersten?= <60147923+kristoffersodersten@users.noreply.github.com> Date: Fri, 4 Sep 2026 07:44:35 +0200 Subject: [PATCH 66/66] fix(runtime): keep smallest model ready after idle Use one explicit generation keep-alive contract across prewarm, runtime requests, launchd, and health reporting. Invalid overrides fail closed and no larger model is selected as fallback. Verified: python3 -m unittest tests.server.test_ollama_client tests.server.test_installation_manager Verified: node --test tests/panel/bootstrap-contract.test.mjs Verified: bash -n scripts/bootstrap-macos-local.sh scripts/install-macos-local.sh installer/macos/scripts/postinstall --- docs/model-runtime-decisions.md | 6 ++--- .../com.selectpilot.bridge.plist.template | 1 + installer/macos/scripts/postinstall | 1 + launchd/com.chromeai.nano.plist | 1 + scripts/bootstrap-macos-local.sh | 13 +++++++--- scripts/install-macos-local.sh | 16 +++++++++--- server/installation_manager.py | 10 +++++++- server/ollama_client.py | 25 +++++++++++++++++++ tests/panel/bootstrap-contract.test.mjs | 5 ++++ tests/server/test_installation_manager.py | 10 ++++++++ tests/server/test_ollama_client.py | 21 +++++++++++++++- 11 files changed, 97 insertions(+), 12 deletions(-) diff --git a/docs/model-runtime-decisions.md b/docs/model-runtime-decisions.md index cf92dfd..931e8a0 100644 --- a/docs/model-runtime-decisions.md +++ b/docs/model-runtime-decisions.md @@ -2,8 +2,8 @@ ## Current production defaults -- Fast / unknown / under 16 GB / Intel: `gemma4:e2b-it-qat`, `num_ctx=16384`. -- Balanced / Apple Silicon with 16 GB or more: `gemma4:e4b-it-qat`, `num_ctx=32768`. +- Fast / automatic default: `gemma4:e2b-it-qat`, `num_ctx=16384`. +- Balanced / manual opt-in: `gemma4:e4b-it-qat`, `num_ctx=32768`. - Advanced: manual opt-in only, `qwen2.5:7b`, `num_ctx=32768`. - Embeddings: `nomic-embed-text-v2-moe:latest`. - Structured generation: `temperature=0` and `seed=42` for every production profile. @@ -14,7 +14,7 @@ Selections above 16,000 characters are rejected before inference with a clear, l Explicit environment overrides remain authoritative and must fail visibly when invalid or unavailable. SelectPilot never substitutes a different installed model, downloads a model without consent, or routes generation to a cloud model. -Installation prewarms the selected generation model with the same context window and deterministic sampling options used at runtime. A failed prewarm is an explicit installation failure; SelectPilot does not report the runtime as ready and does not silently choose another model. `CHROMEAI_OLLAMA_SEED` may override the seed for controlled verification, but invalid or negative values fail closed. +Installation prewarms the selected generation model with the same context window, deterministic sampling options, and keep-alive contract used at runtime. Generation requests set `keep_alive=-1` so Ollama keeps the smallest qualified model resident after idle instead of reloading it inside the interactive request timeout. The explicit `CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS` override accepts `-1` or a non-negative integer number of seconds and fails closed on invalid values. A failed prewarm is an explicit installation failure; SelectPilot does not report the runtime as ready and does not silently choose another model. `CHROMEAI_OLLAMA_SEED` may override the seed for controlled verification, but invalid or negative values fail closed. ## Deferred decisions diff --git a/installer/macos/com.selectpilot.bridge.plist.template b/installer/macos/com.selectpilot.bridge.plist.template index 11069e8..3f827e9 100644 --- a/installer/macos/com.selectpilot.bridge.plist.template +++ b/installer/macos/com.selectpilot.bridge.plist.template @@ -29,6 +29,7 @@ CHROMEAI_OLLAMA_FAST_NUM_CTX__OLLAMA_FAST_NUM_CTX__ CHROMEAI_MAX_INPUT_CHARS__MAX_INPUT_CHARS__ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ + CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS__OLLAMA_KEEP_ALIVE_SECONDS__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_RUNTIME_HASH__RUNTIME_HASH__ CHROMEAI_AUTO_PULL_MODELS0 diff --git a/installer/macos/scripts/postinstall b/installer/macos/scripts/postinstall index 1164281..87ed0b6 100755 --- a/installer/macos/scripts/postinstall +++ b/installer/macos/scripts/postinstall @@ -75,6 +75,7 @@ sed \ -e "s|__OLLAMA_FAST_NUM_CTX__|$OLLAMA_FAST_NUM_CTX|g" \ -e "s|__MAX_INPUT_CHARS__|$MAX_INPUT_CHARS|g" \ -e "s|__OLLAMA_SEED__|42|g" \ + -e "s|__OLLAMA_KEEP_ALIVE_SECONDS__|-1|g" \ -e "s|__ENTITLEMENT_AUTHORITY_URL__|https://license.selectpilot.app|g" \ "$SOURCE_ROOT/launchd/com.selectpilot.bridge.plist.template" > "$AGENT_PATH" chown "$CONSOLE_USER" "$AGENT_PATH" diff --git a/launchd/com.chromeai.nano.plist b/launchd/com.chromeai.nano.plist index df86fcc..8261b3d 100644 --- a/launchd/com.chromeai.nano.plist +++ b/launchd/com.chromeai.nano.plist @@ -32,6 +32,7 @@ CHROMEAI_OLLAMA_FAST_NUM_CTX__OLLAMA_FAST_NUM_CTX__ CHROMEAI_MAX_INPUT_CHARS__MAX_INPUT_CHARS__ CHROMEAI_OLLAMA_SEED__OLLAMA_SEED__ + CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS__OLLAMA_KEEP_ALIVE_SECONDS__ CHROMEAI_BINARY_HASH__BINARY_HASH__ CHROMEAI_RUNTIME_HASH__RUNTIME_HASH__ CHROMEAI_AUTO_PULL_MODELS0 diff --git a/scripts/bootstrap-macos-local.sh b/scripts/bootstrap-macos-local.sh index 53bb91e..fde438f 100755 --- a/scripts/bootstrap-macos-local.sh +++ b/scripts/bootstrap-macos-local.sh @@ -7,6 +7,7 @@ SKIP_OLLAMA_INSTALL="0" SKIP_MODEL_PULL="0" PLAN_ONLY="0" BRIDGE_HEALTH_URL="http://127.0.0.1:8083/health" +GENERATION_KEEP_ALIVE_SECONDS="${CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS:--1}" STATUS_OLLAMA_INSTALL="pending" STATUS_OLLAMA_RUNNING="pending" @@ -41,7 +42,7 @@ while [[ $# -gt 0 ]]; do done read_profile_json() { - python3 - "$PROFILE" "$ROOT" <<'PY' + python3 - "$PROFILE" "$ROOT" "$GENERATION_KEEP_ALIVE_SECONDS" <<'PY' import json import sys from pathlib import Path @@ -50,6 +51,7 @@ profile = sys.argv[1] root = Path(sys.argv[2]) sys.path.insert(0, str(root / "server")) +from ollama_client import parse_generation_keep_alive_seconds from runtime_profiles import build_bootstrap_commands, get_runtime_profile, recommend_runtime_profile recommendation = recommend_runtime_profile() @@ -57,6 +59,7 @@ selected = recommendation["recommended_profile"] if profile == "auto" else profi runtime_profile = get_runtime_profile(selected) commands = build_bootstrap_commands(runtime_profile.key, root) reason = recommendation["reason"] if profile == "auto" else "Explicit profile selected by operator." +generation_keep_alive_seconds = parse_generation_keep_alive_seconds(sys.argv[3]) print(json.dumps({ "selected_profile": runtime_profile.key, @@ -68,6 +71,7 @@ print(json.dumps({ "num_ctx": runtime_profile.num_ctx, "fast_num_ctx": runtime_profile.fast_num_ctx, "max_input_chars": runtime_profile.max_input_chars, + "generation_keep_alive_seconds": generation_keep_alive_seconds, "generation_routes": commands["generation_routes"], "command": commands["command"], })) @@ -84,6 +88,7 @@ payload = json.loads(sys.argv[1]) for key in ( "selected_profile", "generation_model", "fast_generation_model", "embedding_model", "num_ctx", "fast_num_ctx", "max_input_chars", "reason", + "generation_keep_alive_seconds", ): print(f"{key.upper()}={shlex.quote(str(payload[key]))}") PY @@ -160,12 +165,13 @@ else fi echo "Preparing local model bundle" -python3 - "$GENERATION_MODEL" "$NUM_CTX" "$FAST_GENERATION_MODEL" "$FAST_NUM_CTX" <<'PY' +python3 - "$GENERATION_MODEL" "$NUM_CTX" "$FAST_GENERATION_MODEL" "$FAST_NUM_CTX" "$GENERATION_KEEP_ALIVE_SECONDS" <<'PY' import json import sys from urllib.request import Request, urlopen routes = [(sys.argv[1], int(sys.argv[2])), (sys.argv[3], int(sys.argv[4]))] +keep_alive_seconds = int(sys.argv[5]) seen = set() for model, num_ctx in reversed(routes): if model in seen: @@ -177,7 +183,7 @@ for model, num_ctx in reversed(routes): "model": model, "prompt": "Return only: ready", "stream": False, - "keep_alive": "10m", + "keep_alive": keep_alive_seconds, "options": { "temperature": 0, "seed": 42, @@ -202,6 +208,7 @@ CHROMEAI_OLLAMA_NUM_CTX="$NUM_CTX" \ CHROMEAI_OLLAMA_FAST_NUM_CTX="$FAST_NUM_CTX" \ CHROMEAI_MAX_INPUT_CHARS="$MAX_INPUT_CHARS" \ CHROMEAI_OLLAMA_SEED=42 \ +CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS="$GENERATION_KEEP_ALIVE_SECONDS" \ "$ROOT/scripts/install-macos-local.sh" STATUS_LAUNCHAGENT="ok" diff --git a/scripts/install-macos-local.sh b/scripts/install-macos-local.sh index 6e315bf..a7ba28a 100755 --- a/scripts/install-macos-local.sh +++ b/scripts/install-macos-local.sh @@ -22,6 +22,7 @@ OLLAMA_NUM_CTX="${CHROMEAI_OLLAMA_NUM_CTX:-16384}" OLLAMA_FAST_NUM_CTX="${CHROMEAI_OLLAMA_FAST_NUM_CTX:-$OLLAMA_NUM_CTX}" MAX_INPUT_CHARS="${CHROMEAI_MAX_INPUT_CHARS:-16000}" OLLAMA_SEED="${CHROMEAI_OLLAMA_SEED:-42}" +OLLAMA_KEEP_ALIVE_SECONDS="${CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS:--1}" RUN_DIR="${CHROMEAI_RUN_DIR:-${HOME}/Library/Application Support/SelectPilot/run}" STATE_DIR="${CHROMEAI_RUNTIME_STATE_DIR:-${HOME}/Library/Application Support/SelectPilot/state}" LOG_DIR="${CHROMEAI_LOG_DIR:-${HOME}/Library/Logs/SelectPilot}" @@ -32,6 +33,11 @@ if ! "$PYTHON_BIN" -c 'import sys; raise SystemExit(0 if sys.version_info >= (3, exit 1 fi +if [[ ! "$OLLAMA_KEEP_ALIVE_SECONDS" =~ ^-1$|^[0-9]+$ ]]; then + echo "CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS must be -1 or a non-negative integer in seconds." >&2 + exit 1 +fi + mkdir -p "${HOME}/Library/LaunchAgents" "$INSTALL_DIR" "$APP_DIR/presets" "$APP_DIR/runtime" mkdir -p "$RUN_DIR" "$STATE_DIR" "$LOG_DIR" @@ -73,6 +79,7 @@ sed \ -e "s|__OLLAMA_FAST_NUM_CTX__|$OLLAMA_FAST_NUM_CTX|g" \ -e "s|__MAX_INPUT_CHARS__|$MAX_INPUT_CHARS|g" \ -e "s|__OLLAMA_SEED__|$OLLAMA_SEED|g" \ + -e "s|__OLLAMA_KEEP_ALIVE_SECONDS__|$OLLAMA_KEEP_ALIVE_SECONDS|g" \ "$TEMPLATE" > "$DEST" launchctl unload "$DEST" 2>/dev/null || true @@ -93,8 +100,9 @@ Next steps: 5. Local bridge URL: http://127.0.0.1:8083 6. Ollama context window: $OLLAMA_NUM_CTX 7. Deterministic seed: $OLLAMA_SEED - 8. Run dir: $RUN_DIR - 9. Runtime state dir: $STATE_DIR - 10. Log dir: $LOG_DIR - 11. Run 'pnpm benchmark:local' to validate latency on this machine. + 8. Generation keep-alive seconds: $OLLAMA_KEEP_ALIVE_SECONDS + 9. Run dir: $RUN_DIR + 10. Runtime state dir: $STATE_DIR + 11. Log dir: $LOG_DIR + 12. Run 'pnpm benchmark:local' to validate latency on this machine. EOF diff --git a/server/installation_manager.py b/server/installation_manager.py index 9688f0a..96355bc 100644 --- a/server/installation_manager.py +++ b/server/installation_manager.py @@ -12,6 +12,11 @@ from pathlib import Path from urllib.request import Request, urlopen +from ollama_client import ( + DEFAULT_GENERATION_KEEP_ALIVE_SECONDS, + GENERATION_KEEP_ALIVE_ENV, + parse_generation_keep_alive_seconds, +) from runtime_profiles import get_runtime_profile, recommend_runtime_profile, required_generation_models OLLAMA_DOWNLOAD_URL = "https://ollama.com/download/Ollama-darwin.zip" @@ -130,13 +135,16 @@ def _warm_model(self, model: str, num_ctx: int) -> None: seed = int(os.environ.get("CHROMEAI_OLLAMA_SEED", "42")) if seed < 0: raise RuntimeError("model_warmup_failed") + keep_alive_seconds = parse_generation_keep_alive_seconds( + os.environ.get(GENERATION_KEEP_ALIVE_ENV, DEFAULT_GENERATION_KEEP_ALIVE_SECONDS), + ) request = Request( "http://127.0.0.1:11434/api/generate", data=json.dumps({ "model": model, "prompt": "Return only: ready", "stream": False, - "keep_alive": "10m", + "keep_alive": keep_alive_seconds, "options": { "temperature": 0, "seed": seed, diff --git a/server/ollama_client.py b/server/ollama_client.py index 3b710f3..55cb85f 100644 --- a/server/ollama_client.py +++ b/server/ollama_client.py @@ -92,6 +92,9 @@ def _build_markdown(summary: str, bullets: list[str], action_items: list[str] | "mxbai-embed-large", ] +DEFAULT_GENERATION_KEEP_ALIVE_SECONDS = -1 +GENERATION_KEEP_ALIVE_ENV = "CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS" + @dataclass(frozen=True) class OllamaConfig: @@ -101,6 +104,7 @@ class OllamaConfig: timeout_seconds: float num_ctx: int seed: int + generation_keep_alive_seconds: int = DEFAULT_GENERATION_KEEP_ALIVE_SECONDS fast_model: str | None = None fast_num_ctx: int | None = None max_input_chars: int = 16000 @@ -130,6 +134,19 @@ def _nonnegative_int(value: Any, name: str) -> int: return parsed +def parse_generation_keep_alive_seconds( + value: Any, + name: str = GENERATION_KEEP_ALIVE_ENV, +) -> int: + try: + parsed = int(value) + except (TypeError, ValueError) as exc: + raise OllamaError(f"{name} must be -1 or a non-negative integer in seconds") from exc + if parsed < -1: + raise OllamaError(f"{name} must be -1 or a non-negative integer in seconds") + return parsed + + class OllamaClient: def __init__(self, config: OllamaConfig | None = None): if config is None: @@ -159,6 +176,9 @@ def __init__(self, config: OllamaConfig | None = None): timeout_seconds=float(os.environ.get("CHROMEAI_OLLAMA_TIMEOUT_SECONDS", "30")), num_ctx=_positive_int(os.environ.get("CHROMEAI_OLLAMA_NUM_CTX", default_num_ctx), "CHROMEAI_OLLAMA_NUM_CTX"), seed=_nonnegative_int(os.environ.get("CHROMEAI_OLLAMA_SEED", "42"), "CHROMEAI_OLLAMA_SEED"), + generation_keep_alive_seconds=parse_generation_keep_alive_seconds( + os.environ.get(GENERATION_KEEP_ALIVE_ENV, DEFAULT_GENERATION_KEEP_ALIVE_SECONDS), + ), fast_model=fast_model, fast_num_ctx=_positive_int( os.environ.get("CHROMEAI_OLLAMA_FAST_NUM_CTX", profile.fast_num_ctx), @@ -360,6 +380,7 @@ def health(self) -> dict[str, Any]: "timeout_seconds": self.config.timeout_seconds, "num_ctx": self.config.num_ctx, "seed": self.config.seed, + "generation_keep_alive_seconds": self.config.generation_keep_alive_seconds, "reachable": False, "status": "degraded", "error": str(e), @@ -402,6 +423,7 @@ def health(self) -> dict[str, Any]: "timeout_seconds": self.config.timeout_seconds, "num_ctx": self.config.num_ctx, "seed": self.config.seed, + "generation_keep_alive_seconds": self.config.generation_keep_alive_seconds, "max_input_chars": self.config.max_input_chars, "task_routes": routes, } @@ -453,6 +475,7 @@ def summarize(self, text: str, title: str | None = None, url: str | None = None, "prompt": prompt, "system": "You write precise summaries for selected text in a browser side panel.", "stream": False, + "keep_alive": self.config.generation_keep_alive_seconds, "format": schema, "options": self._generation_options(0.0, route["num_ctx"]), } @@ -510,6 +533,7 @@ def agent(self, prompt: str, context: dict[str, Any] | None = None) -> dict[str, "prompt": user_prompt, "system": "You are a practical browser copilot that rewrites and structures selected text locally.", "stream": False, + "keep_alive": self.config.generation_keep_alive_seconds, "format": schema, "options": self._generation_options(0.0, route["num_ctx"]), } @@ -611,6 +635,7 @@ def extract( "prompt": prompt, "system": "You generate clean structured extraction results for highlighted browser text.", "stream": False, + "keep_alive": self.config.generation_keep_alive_seconds, "format": preset.schema, "options": self._generation_options(0.0, route["num_ctx"]), } diff --git a/tests/panel/bootstrap-contract.test.mjs b/tests/panel/bootstrap-contract.test.mjs index 6fc8bfa..6c664d6 100644 --- a/tests/panel/bootstrap-contract.test.mjs +++ b/tests/panel/bootstrap-contract.test.mjs @@ -23,6 +23,7 @@ test('bootstrap plan exposes exact fail-closed Fast runtime contract without sid assert.equal(plan.num_ctx, 16_384); assert.equal(plan.fast_num_ctx, 16_384); assert.equal(plan.max_input_chars, 16_000); + assert.equal(plan.generation_keep_alive_seconds, -1); assert.equal(plan.reason, 'Explicit profile selected by operator.'); }); @@ -76,6 +77,9 @@ test('macOS service installs outside the source checkout before launchd executio assert.match(launchAgent, /__INSTALLED_BINARY__/); assert.match(launchAgent, /__INSTALL_DIR__/); assert.match(launchAgent, /CHROMEAI_RUNTIME_HASH/); + assert.match(installer, /CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS/); + assert.match(installer, /must be -1 or a non-negative integer in seconds/); + assert.match(launchAgent, /CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS/); assert.doesNotMatch(launchAgent, /__PROJECT_ROOT__/); }); @@ -118,6 +122,7 @@ test('production macOS package preserves explicit task routing and profile truth '__OLLAMA_NUM_CTX__', '__OLLAMA_FAST_NUM_CTX__', '__MAX_INPUT_CHARS__', + '__OLLAMA_KEEP_ALIVE_SECONDS__', ]) { assert.ok(postinstall.includes(placeholder), `postinstall does not bind ${placeholder}`); assert.ok(launchAgent.includes(placeholder), `LaunchAgent does not expose ${placeholder}`); diff --git a/tests/server/test_installation_manager.py b/tests/server/test_installation_manager.py index 9db7dae..fcd101b 100644 --- a/tests/server/test_installation_manager.py +++ b/tests/server/test_installation_manager.py @@ -45,6 +45,7 @@ def test_warmup_uses_deterministic_profile_options(self) -> None: request = mocked.call_args.args[0] payload = json.loads(request.data) self.assertEqual(payload["model"], "gemma4:e4b-it-qat") + self.assertEqual(payload["keep_alive"], -1) self.assertEqual(payload["options"], { "temperature": 0, "seed": 42, @@ -65,6 +66,15 @@ def test_warmup_uses_runtime_seed_override(self) -> None: payload = json.loads(mocked.call_args.args[0].data) self.assertEqual(payload["options"]["seed"], 7) + def test_warmup_rejects_invalid_keep_alive_override(self) -> None: + with patch.dict( + "installation_manager.os.environ", + {"CHROMEAI_OLLAMA_KEEP_ALIVE_SECONDS": "forever"}, + clear=True, + ), patch("installation_manager.urlopen"): + with self.assertRaisesRegex(RuntimeError, "-1 or a non-negative integer"): + InstallationManager()._warm_model("gemma4:e2b-it-qat", 16384) + if __name__ == "__main__": unittest.main() diff --git a/tests/server/test_ollama_client.py b/tests/server/test_ollama_client.py index 6c6e1c3..69f9c95 100644 --- a/tests/server/test_ollama_client.py +++ b/tests/server/test_ollama_client.py @@ -12,7 +12,12 @@ if str(SERVER_DIR) not in sys.path: sys.path.insert(0, str(SERVER_DIR)) -from ollama_client import OllamaClient, OllamaConfig, OllamaError # noqa: E402 +from ollama_client import ( # noqa: E402 + OllamaClient, + OllamaConfig, + OllamaError, + parse_generation_keep_alive_seconds, +) class StubOllamaClient(OllamaClient): @@ -80,6 +85,19 @@ def test_summarize_sends_explicit_context_window(self) -> None: self.assertEqual(client.last_payload["options"]["num_ctx"], 16384) self.assertEqual(client.last_payload["options"]["seed"], 42) self.assertEqual(client.last_payload["options"]["temperature"], 0.0) + self.assertEqual(client.last_payload["keep_alive"], -1) + + def test_generation_keep_alive_override_is_validated(self) -> None: + self.assertEqual(parse_generation_keep_alive_seconds("-1"), -1) + self.assertEqual(parse_generation_keep_alive_seconds("3600"), 3600) + with self.assertRaisesRegex(OllamaError, "-1 or a non-negative integer"): + parse_generation_keep_alive_seconds("forever") + with self.assertRaisesRegex(OllamaError, "-1 or a non-negative integer"): + parse_generation_keep_alive_seconds("-2") + + def test_health_exposes_generation_keep_alive_contract(self) -> None: + client = StubOllamaClient({}) + self.assertEqual(client.health()["generation_keep_alive_seconds"], -1) def test_balanced_routes_structured_and_agent_tasks_to_exact_models(self) -> None: client = StubOllamaClient({ @@ -136,6 +154,7 @@ def test_all_structured_routes_send_closed_object_schemas(self) -> None: client.response = responses.pop(0) self.assertEqual([payload["format"]["type"] for payload in payloads], ["object"] * 3) self.assertTrue(all(payload["format"]["additionalProperties"] is False for payload in payloads)) + self.assertTrue(all(payload["keep_alive"] == -1 for payload in payloads)) if __name__ == "__main__":
    +
    -
    Billing
    -

    Attach paid token

    +
    Access
    +

    Keep your flow available

    - +
    - - + +
    -

    No entitlement token attached yet.

    +
    + + +
    +

    No active access yet.

    @@ -159,7 +131,7 @@

    Turn selected text into structure

    -
    +
    Result
    @@ -178,6 +150,9 @@

    Ready

    Highlight content, then run the primary action to generate structured output locally.
    +
    diff --git a/panel/panel.js b/panel/panel.js index 2968197..b7638ff 100644 --- a/panel/panel.js +++ b/panel/panel.js @@ -20,8 +20,10 @@ const memoryShellEl = $('#memory-shell'); const memoryStatusEl = $('#memory-status'); const memoryMetaEl = $('#memory-meta'); const memoryTargetEl = $('#memory-target'); -const orderIdInputEl = $('#order-id'); -const syncOrderButtonEl = $('#btn-sync-order'); +const licenseTokenInputEl = $('#license-token'); +const attachLicenseButtonEl = $('#btn-attach-license'); +const startTrialButtonEl = $('#btn-start-trial'); +const viewPlansButtonEl = $('#btn-view-plans'); const entitlementStatusEl = $('#entitlement-status'); const memoryToggleButtonEl = $('#btn-memory-toggle'); const memoryInspectButtonEl = $('#btn-memory-inspect'); @@ -61,6 +63,8 @@ const runtimeMetaOperationEl = $('#runtime-meta-operation'); const runtimeMetaStepEl = $('#runtime-meta-step'); const runtimeMetaTraceEl = $('#runtime-meta-trace'); const runtimeMetaEventsEl = $('#runtime-meta-events'); +const resultShellEl = $('#result-shell'); +const processingFieldEl = $('#processing-field'); const actionButtons = Array.from(document.querySelectorAll('.primary-action, .secondary-grid button, .advanced-grid button')); const ENTITLEMENT_FRESH_MS = 15 * 60 * 1000; const FIRST_RUN_COMPLETED_KEY = 'selectpilot_first_run_completed_v1'; @@ -98,9 +102,19 @@ let memorySnapshot = { lastUpdatedAt: null, }; let entitlementSnapshot = null; +let bridgeAvailable = false; +let installationSnapshot = { + state: 'idle', + label: 'Ready to install', + progress: 0, + profile: null, + action_required: null, +}; +let installationPollTimer = null; let firstRunCompleted = false; const BENCHMARK_CACHE_KEY = 'selectpilot_runtime_benchmark_v1'; const RUNTIME_META_MAX_EVENTS = 6; +const XRAY_ENABLED = new URLSearchParams(location.search).has('xray'); let runtimeMetaEventSource = null; let runtimeMetaReconnectTimer = null; let runtimeMetaReconnectDelayMs = 1200; @@ -118,19 +132,16 @@ const runtimeMetaOverlayState = { }; let intentSuggestions = []; let selectedIntentSuggestion = null; -const FAST_INSTALL_COMMANDS = [ - 'ollama pull gemma4:e2b-it-qat', - 'ollama pull nomic-embed-text-v2-moe:latest', - 'CHROMEAI_OLLAMA_MODEL=gemma4:e2b-it-qat CHROMEAI_OLLAMA_NUM_CTX=16384 ./scripts/install-macos-local.sh', -].join('\n'); -const QUICK_SETUP_COMMANDS = [ - 'pnpm setup:local', - 'curl http://127.0.0.1:8083/health', -].join('\n'); function setStatus(text) { + if (resultShellEl?.getAttribute('aria-busy') === 'true') + return; if (statusEl) statusEl.textContent = text; } +function setSilentProcessing(active) { + processingFieldEl?.classList.toggle('is-active', active); + resultShellEl?.setAttribute('aria-busy', String(active)); +} function setStatusBar(text) { if (statusBar) statusBar.textContent = text; @@ -164,13 +175,6 @@ function shorten(text, max = 220) { return trimmed; return `${trimmed.slice(0, max - 1).trimEnd()}…`; } -function getEfficiencyScore(snapshot) { - if (!snapshot) - return null; - const avgLatency = (snapshot.extract_latency_ms + snapshot.summarize_latency_ms) / 2; - const score = Math.round(100 - avgLatency / 40); - return Math.max(0, Math.min(100, score)); -} function formatPrivacyVerifiedAt(iso) { if (!iso) return 'Awaiting proof'; @@ -544,16 +548,6 @@ function getEffectiveRecommendedProfileKey() { function getEffectiveRecommendedProfile() { return getRuntimeProfile(getEffectiveRecommendedProfileKey()); } -function getEffectiveRecommendationReason() { - if (!benchmarkSnapshot) - return runtimeProfilesPayload.reason; - const benchmarkProfile = getRuntimeProfile(benchmarkSnapshot.recommended_profile).label; - const autoProfile = getRuntimeProfile(benchmarkSnapshot.auto_profile || runtimeProfilesPayload.recommended_profile).label; - if (benchmarkSnapshot.recommended_profile === (benchmarkSnapshot.auto_profile || runtimeProfilesPayload.recommended_profile)) { - return `Benchmark confirms the ${benchmarkProfile} profile for this workload.`; - } - return `Benchmark overrides the hardware heuristic: use ${benchmarkProfile} for this workload instead of the auto ${autoProfile} profile.`; -} async function loadBenchmarkSnapshot() { const cached = await getJSON(BENCHMARK_CACHE_KEY); if (!cached || !cached.recommended_profile) @@ -756,12 +750,12 @@ function renderEntitlementStatus() { const tier = entitlementSnapshot?.tier || 'essential'; const token = entitlementSnapshot?.token; if (!token) { - entitlementStatusEl.textContent = 'No entitlement token attached yet.'; + entitlementStatusEl.textContent = 'No active access yet.'; return; } const suffix = formatEntitlementUpdatedAt(entitlementSnapshot?.cachedAt); const cacheState = getEntitlementCacheState(entitlementSnapshot); - entitlementStatusEl.textContent = `Token attached · tier ${tier} (${cacheState}) · ${suffix}`; + entitlementStatusEl.textContent = `${tier} access · ${cacheState} · ${suffix}`; } async function refreshEntitlementStatus() { try { @@ -774,33 +768,40 @@ async function refreshEntitlementStatus() { renderSelectionState(); syncControlAvailability(); } -async function doSyncOrderToken() { - const orderId = orderIdInputEl?.value.trim() || ''; - if (!orderId) - throw new Error('Order ID is required'); - setStatus('Checking order status...'); - const orderRes = await fetch(endpoints.billingOrderStatus(orderId), { cache: 'no-store' }); - if (!orderRes.ok) { - if (orderRes.status === 404) - throw new Error('No payment detected for this order ID'); - throw new Error(`Order lookup failed (${orderRes.status})`); - } - const order = await orderRes.json(); - if (!order.paid || !order.token) { - const confirmations = order.confirmations || 0; - const needed = order.confirmations_required || 0; - throw new Error(`No payment detected yet (${confirmations}/${needed} confirmations)`); - } - if (entitlementSnapshot?.token && entitlementSnapshot.token === order.token) { - await request('entitlement:refresh'); - await refreshEntitlementStatus(); - setStatus('Order already synced; entitlement refreshed'); - return; - } - await request('license:attach_token', { token: order.token }); +async function attachLicenseToken(token = licenseTokenInputEl?.value.trim() || '') { + if (!token) + throw new Error('Enter your license key'); + setStatus('Activating access...'); + await request('license:attach_token', { token }); await request('entitlement:refresh'); await refreshEntitlementStatus(); - setStatus('Token attached and entitlement refreshed'); + if (licenseTokenInputEl) + licenseTokenInputEl.value = ''; + setStatus('Access ready'); +} +async function getInstallationId() { + const key = 'selectpilot_installation_id'; + const stored = await chrome.storage.local.get(key); + if (typeof stored[key] === 'string' && stored[key]) + return stored[key]; + const value = crypto.randomUUID(); + await chrome.storage.local.set({ [key]: value }); + return value; +} +async function startTrial() { + setStatus('Preparing your trial...'); + const response = await fetch(endpoints.licenseTrial, { + method: 'POST', + headers: { 'Content-Type': 'application/json' }, + body: JSON.stringify({ installation_id: await getInstallationId() }), + cache: 'no-store', + }); + if (!response.ok) + throw new Error('Trial could not be started'); + const result = await response.json(); + if (!result.token) + throw new Error('Trial response was incomplete'); + await attachLicenseToken(result.token); } async function refreshMemoryStatus() { try { @@ -853,10 +854,10 @@ function syncControlAvailability() { memoryExportButtonEl.disabled = isBusy || !flowExportSupported || !canExport; if (memoryDeleteButtonEl) memoryDeleteButtonEl.disabled = memoryActionsLocked || memorySnapshot.entries === 0; - if (syncOrderButtonEl) { - const hasOrderId = Boolean(orderIdInputEl?.value.trim()); - syncOrderButtonEl.disabled = isBusy || !hasOrderId; - } + if (attachLicenseButtonEl) + attachLicenseButtonEl.disabled = isBusy || !licenseTokenInputEl?.value.trim(); + if (startTrialButtonEl) + startTrialButtonEl.disabled = isBusy || Boolean(entitlementSnapshot?.token); if (intentExecuteButtonEl) { const hasIntent = Boolean(intentInputEl?.value.trim()); intentExecuteButtonEl.disabled = isBusy || !runtimeReady || !selectionReady || !hasIntent; @@ -884,175 +885,84 @@ function syncPresetHelp() { if (extractHelpEl) extractHelpEl.textContent = preset.description; } -function buildMetric(label, value) { - const metric = document.createElement('div'); - metric.className = 'runtime-metric'; - const kicker = document.createElement('span'); - kicker.className = 'section-kicker'; - kicker.textContent = label; - const strong = document.createElement('strong'); - strong.textContent = value; - metric.append(kicker, strong); - return metric; -} -function createProfileCard(profile, recommendedKey, copyLabel = 'Copy command') { - const card = document.createElement('div'); - card.className = 'profile-card'; - if (profile.key === recommendedKey) - card.classList.add('is-recommended'); - const header = document.createElement('div'); - header.className = 'profile-header'; - const title = document.createElement('div'); - title.className = 'profile-title'; - title.textContent = profile.label; - header.append(title); - if (profile.key === recommendedKey) { - const badge = document.createElement('span'); - badge.className = 'profile-badge'; - badge.textContent = 'Recommended'; - header.append(badge); - } - const description = document.createElement('p'); - description.className = 'selection-copy'; - description.textContent = profile.description; - const stack = document.createElement('div'); - stack.className = 'profile-stack'; - stack.append(buildMetric('Generation', profile.generation_model), buildMetric('Embedding', profile.embedding_model), buildMetric('Latency', profile.target_latency)); - const intendedFor = document.createElement('p'); - intendedFor.className = 'selection-copy'; - intendedFor.textContent = profile.intended_for; - const actions = document.createElement('div'); - actions.className = 'runtime-actions'; - const copyButton = document.createElement('button'); - copyButton.type = 'button'; - copyButton.textContent = copyLabel; - copyButton.addEventListener('click', async () => { - await navigator.clipboard.writeText(profile.command); - setStatus(`${profile.label} bootstrap command copied`); - }); - actions.append(copyButton); - card.append(header, description, stack, intendedFor, actions); - return card; -} function renderRuntimeState() { clearNode(runtimeStateEl); if (!runtimeStateEl) return; runtimeStateEl.classList.add('is-visible'); - const profilesGrid = document.createElement('div'); - profilesGrid.className = 'profile-grid'; - const effectiveRecommendedProfileKey = getEffectiveRecommendedProfileKey(); - for (const profile of runtimeProfilesPayload.profiles) { - profilesGrid.append(createProfileCard(profile, effectiveRecommendedProfileKey)); - } if (runtimeSnapshot.ok) { - const wrapper = document.createElement('div'); - wrapper.className = 'runtime-grid'; - wrapper.append(buildMetric('Execution', 'Local-only ready'), buildMetric('Provider', 'Ollama on-device'), buildMetric('Ignored remote', `${runtimeSnapshot.ignoredRemoteCount} models`)); - const copy = document.createElement('p'); - copy.className = 'runtime-copy'; - copy.textContent = `Runtime ready. ${runtimeSnapshot.activeModel} is active locally, and the selected-text path stays inside the local bridge.`; - const benchmarkBlock = document.createElement('div'); - benchmarkBlock.className = 'benchmark-block'; - const benchmarkHeader = document.createElement('div'); - benchmarkHeader.className = 'profile-header'; - const benchmarkTitle = document.createElement('div'); - benchmarkTitle.className = 'profile-title'; - benchmarkTitle.textContent = 'Profile fit'; - benchmarkHeader.append(benchmarkTitle); - benchmarkBlock.append(benchmarkHeader); - const benchmarkCopy = document.createElement('p'); - benchmarkCopy.className = 'selection-copy'; - if (benchmarkSnapshot) { - benchmarkCopy.textContent = getEffectiveRecommendationReason(); - } - else { - benchmarkCopy.textContent = 'Run a local benchmark to confirm that the current runtime matches the smallest viable profile for this machine.'; - } - benchmarkBlock.append(benchmarkCopy); - const benchmarkMetrics = document.createElement('div'); - benchmarkMetrics.className = 'runtime-grid'; - const efficiencyScore = getEfficiencyScore(benchmarkSnapshot); - if (benchmarkSnapshot) { - benchmarkMetrics.append(buildMetric('Extract JSON', `${benchmarkSnapshot.extract_latency_ms} ms`), buildMetric('Summarize', `${benchmarkSnapshot.summarize_latency_ms} ms`), buildMetric('Efficiency', `${efficiencyScore}/100`), buildMetric('Recommended', getEffectiveRecommendedProfile().label)); - } - else { - benchmarkMetrics.append(buildMetric('Extract JSON', 'Pending'), buildMetric('Summarize', 'Pending'), buildMetric('Efficiency', 'Pending'), buildMetric('Recommended', getEffectiveRecommendedProfile().label)); - } - benchmarkBlock.append(benchmarkMetrics); - const benchmarkActions = document.createElement('div'); - benchmarkActions.className = 'runtime-actions'; - const benchmarkButton = document.createElement('button'); - benchmarkButton.type = 'button'; - benchmarkButton.id = 'btn-run-benchmark'; - benchmarkButton.textContent = benchmarkSnapshot ? 'Run benchmark again' : 'Run benchmark'; - benchmarkActions.append(benchmarkButton); - benchmarkBlock.append(benchmarkActions); - runtimeStateEl.append(wrapper, copy, benchmarkBlock, profilesGrid); + runtimeStateEl.classList.remove('is-visible'); return; } - const header = document.createElement('div'); - header.className = 'output-card'; - header.innerHTML = '
    Human intervention

    Local runtime not ready

    '; - const copy = document.createElement('p'); - copy.className = 'runtime-copy'; - copy.textContent = - 'SelectPilot runs only through local Ollama. Follow these three tiny steps and press re-check when done.'; - const tutorialSteps = document.createElement('div'); - tutorialSteps.className = 'tutorial-steps'; - const stepInstall = document.createElement('div'); - stepInstall.className = 'tutorial-step'; - stepInstall.innerHTML = 'Step 1Copy and run the one-command setup in Terminal.'; - const stepLoad = document.createElement('div'); - stepLoad.className = 'tutorial-step'; - stepLoad.innerHTML = 'Step 2Open chrome://extensions, enable Developer Mode, then Load unpacked.'; - const stepVerify = document.createElement('div'); - stepVerify.className = 'tutorial-step'; - stepVerify.innerHTML = 'Step 3Highlight any text, click Extract JSON, and confirm output appears.'; - tutorialSteps.append(stepInstall, stepLoad, stepVerify); - const wrapper = document.createElement('div'); - wrapper.className = 'runtime-grid'; - const recommended = getEffectiveRecommendedProfile(); - wrapper.append(buildMetric('Recommended profile', recommended.label), buildMetric('Generation model', recommended.generation_model), buildMetric('Embedding model', recommended.embedding_model)); - const actions = document.createElement('div'); - actions.className = 'runtime-actions'; - const copySetup = document.createElement('button'); - copySetup.type = 'button'; - copySetup.textContent = 'Copy one-command setup'; - copySetup.addEventListener('click', async () => { - await navigator.clipboard.writeText(QUICK_SETUP_COMMANDS); - setStatus('One-command setup copied'); - }); - actions.append(copySetup); - const copyAdvancedSetup = document.createElement('button'); - copyAdvancedSetup.type = 'button'; - copyAdvancedSetup.textContent = 'Copy manual fallback'; - copyAdvancedSetup.addEventListener('click', async () => { - await navigator.clipboard.writeText(recommended.command || FAST_INSTALL_COMMANDS); - setStatus('Fallback setup copied'); - }); - actions.append(copyAdvancedSetup); - const refresh = document.createElement('button'); - refresh.type = 'button'; - refresh.textContent = 'Re-check runtime'; - refresh.addEventListener('click', () => { - void refreshRuntime(); - }); - actions.append(refresh); - if (runtimeSnapshot.error || runtimeSnapshot.hint) { - const note = document.createElement('p'); - note.className = 'runtime-copy'; - note.textContent = runtimeSnapshot.error || runtimeSnapshot.hint || ''; - const reason = document.createElement('p'); - reason.className = 'runtime-copy'; - reason.textContent = getEffectiveRecommendationReason(); - runtimeStateEl.append(header, copy, wrapper, tutorialSteps, actions, reason, profilesGrid, note); + if (!bridgeAvailable) { + const helper = createCard('Installation', 'One small helper is needed', 'It keeps selected text on this computer and prepares the right local model.'); + const link = document.createElement('a'); + link.className = 'primary-action'; + link.href = 'https://selectpilot.app/downloads/SelectPilot-Installer.pkg'; + link.target = '_blank'; + link.rel = 'noreferrer'; + link.textContent = 'Install local helper'; + runtimeStateEl.append(helper, link); return; } - const reason = document.createElement('p'); - reason.className = 'runtime-copy'; - reason.textContent = getEffectiveRecommendationReason(); - runtimeStateEl.append(header, copy, wrapper, tutorialSteps, actions, reason, profilesGrid); + const installation = createCard('Installation', installationSnapshot.label, 'The right local setup will be chosen automatically for this Mac.'); + const progress = document.createElement('div'); + progress.className = 'installation-progress'; + progress.setAttribute('role', 'progressbar'); + progress.setAttribute('aria-valuemin', '0'); + progress.setAttribute('aria-valuemax', '100'); + progress.setAttribute('aria-valuenow', String(installationSnapshot.progress)); + const fill = document.createElement('span'); + fill.style.width = `${installationSnapshot.progress}%`; + progress.append(fill); + installation.append(progress); + if (installationSnapshot.state === 'installing') { + runtimeStateEl.append(installation); + scheduleInstallationPoll(); + return; + } + const action = document.createElement('button'); + action.className = 'primary-action'; + action.type = 'button'; + action.textContent = installationSnapshot.state === 'action_required' ? 'Try again' : 'Install Ollama and continue'; + action.addEventListener('click', () => void startLocalInstallation()); + runtimeStateEl.append(installation, action); + return; +} +async function refreshInstallationStatus() { + const response = await fetch(endpoints.installationStatus, { cache: 'no-store' }); + if (!response.ok) + throw new Error('installation_status_unavailable'); + installationSnapshot = await response.json(); +} +async function startLocalInstallation() { + const response = await fetch(endpoints.installationStart, { + method: 'POST', + headers: { 'Content-Type': 'application/json' }, + body: JSON.stringify({ consent: true }), + cache: 'no-store', + }); + if (!response.ok) + throw new Error('Installation could not start'); + installationSnapshot = await response.json(); + renderRuntimeState(); +} +function scheduleInstallationPoll() { + if (installationPollTimer !== null) + return; + installationPollTimer = window.setTimeout(async () => { + installationPollTimer = null; + try { + await refreshInstallationStatus(); + if (installationSnapshot.state === 'ready') + await refreshRuntime(); + else + renderRuntimeState(); + } + catch { + renderRuntimeState(); + } + }, 900); } function renderSelectionState() { clearNode(selectionCardEl); @@ -1108,7 +1018,11 @@ async function refreshSelectionPreview() { title: preview.title || '', url: preview.url || '', hasSelection: Boolean(preview.hasSelection), + pageColor: typeof preview.pageColor === 'string' ? preview.pageColor : '', }; + if (/^(?:rgb|hsl)a?\([^)]*\)$|^#[0-9a-f]{6}$/i.test(selectionPreview.pageColor || '')) { + document.documentElement.style.setProperty('--context-tint', selectionPreview.pageColor || 'transparent'); + } setSelectionContext({ selectionOrigin: selectionPreview.hasSelection ? 'selection' : 'page_context', contentLength: selectionPreview.hasSelection ? selectionPreview.selection.length : selectionPreview.pageText.length, @@ -1137,6 +1051,13 @@ async function refreshRuntime() { }; } const health = await fetchHealth(); + bridgeAvailable = true; + try { + await refreshInstallationStatus(); + } + catch { + installationSnapshot = { state: 'idle', label: 'Ready to install', progress: 0, profile: null, action_required: null }; + } privacyProofSnapshot = await fetchPrivacyProof(); runtimeSnapshot = { ok: Boolean(health?.ok), @@ -1179,6 +1100,7 @@ async function refreshRuntime() { renderRuntimeState(); } catch (e) { + bridgeAvailable = false; runtimeProfilesPayload = { profiles: RUNTIME_PROFILES, recommended_profile: 'fast', @@ -1415,19 +1337,23 @@ async function doMemoryDelete() { setStatus('Memory ledger deleted'); } function bindActions() { - const wrap = (fn) => async () => { + const wrap = (fn, mode = 'explicit') => async () => { isBusy = true; + if (mode === 'silent') + setSilentProcessing(true); syncControlAvailability(); try { await fn(); } catch (e) { + setSilentProcessing(false); const text = formatPanelError(e); setStatus(text); setStatusBar(text); } finally { isBusy = false; + setSilentProcessing(false); syncControlAvailability(); void Promise.all([refreshSelectionPreview(), refreshMemoryStatus(), refreshEntitlementStatus()]); } @@ -1435,21 +1361,21 @@ function bindActions() { refreshButtonEl?.addEventListener('click', () => { void Promise.all([refreshRuntime(), refreshSelectionPreview(), refreshMemoryStatus(), refreshEntitlementStatus()]); }); - $('#btn-extract')?.addEventListener('click', wrap(() => doExtract())); + $('#btn-extract')?.addEventListener('click', wrap(() => doExtract(), 'silent')); selectionCardEl?.addEventListener('click', (event) => { const target = event.target; if (target?.id === 'btn-first-run-example') - void wrap(() => doFirstRunExample())(); + void wrap(() => doFirstRunExample(), 'silent')(); }); - $('#btn-summarize')?.addEventListener('click', wrap(() => doSummarize())); - $('#btn-rewrite')?.addEventListener('click', wrap(() => doRewrite())); - $('#btn-actions')?.addEventListener('click', wrap(() => doActions())); - $('#btn-ask')?.addEventListener('click', wrap(() => doAsk())); + $('#btn-summarize')?.addEventListener('click', wrap(() => doSummarize(), 'silent')); + $('#btn-rewrite')?.addEventListener('click', wrap(() => doRewrite(), 'silent')); + $('#btn-actions')?.addEventListener('click', wrap(() => doActions(), 'silent')); + $('#btn-ask')?.addEventListener('click', wrap(() => doAsk(), 'silent')); memoryToggleButtonEl?.addEventListener('click', wrap(() => doMemoryToggle())); memoryInspectButtonEl?.addEventListener('click', wrap(() => doMemoryInspect())); memoryExportButtonEl?.addEventListener('click', wrap(() => doMemoryExport())); memoryDeleteButtonEl?.addEventListener('click', wrap(() => doMemoryDelete())); - intentExecuteButtonEl?.addEventListener('click', wrap(() => doExecuteIntent())); + intentExecuteButtonEl?.addEventListener('click', wrap(() => doExecuteIntent(), 'silent')); intentClearButtonEl?.addEventListener('click', () => clearIntentInput()); intentInputEl?.addEventListener('input', () => { selectedIntentSuggestion = null; @@ -1459,10 +1385,14 @@ function bindActions() { if (event.key !== 'Enter') return; event.preventDefault(); - void wrap(() => doExecuteIntent())(); + void wrap(() => doExecuteIntent(), 'silent')(); }); - syncOrderButtonEl?.addEventListener('click', wrap(() => doSyncOrderToken())); - orderIdInputEl?.addEventListener('input', () => syncControlAvailability()); + attachLicenseButtonEl?.addEventListener('click', wrap(() => attachLicenseToken())); + startTrialButtonEl?.addEventListener('click', wrap(() => startTrial())); + viewPlansButtonEl?.addEventListener('click', wrap(async () => { + await chrome.tabs.create({ url: 'https://selectpilot.app/pricing', active: true }); + })); + licenseTokenInputEl?.addEventListener('input', () => syncControlAvailability()); extractPresetEl?.addEventListener('change', () => syncPresetHelp()); runtimeStateEl?.addEventListener('click', (event) => { const target = event.target; @@ -1491,6 +1421,14 @@ function bindActions() { populatePresetOptions(); bindActions(); async function initialize() { + const storedSettings = await chrome.storage.sync.get('selectpilot_settings'); + const settings = (storedSettings.selectpilot_settings || {}); + const textSize = typeof settings.textSize === 'string' ? settings.textSize : 'standard'; + document.body.dataset.controlSide = settings.controlSide === 'left' ? 'left' : 'right'; + document.body.dataset.textSize = ['large', 'larger'].includes(textSize) ? textSize : 'standard'; + document.body.dataset.highContrast = settings.highContrast ? 'true' : 'false'; + document.body.dataset.reduceMotion = settings.reduceMotion ? 'true' : 'false'; + currentResultView = settings.resultView === 'structured' ? 'structured' : 'readable'; const topologyValidation = validateTopologyMap(); const requiredTopologyComponents = [ 'panel_header', @@ -1515,8 +1453,10 @@ async function initialize() { } setVisiblePanels(['selection_surface', 'runtime_surface', 'report_surface']); firstRunCompleted = Boolean(await getJSON(FIRST_RUN_COMPLETED_KEY)); - renderRuntimeMetaOverlay(); - void connectRuntimeMetaStream(); + if (XRAY_ENABLED) { + renderRuntimeMetaOverlay(); + void connectRuntimeMetaStream(); + } refreshIntentSuggestions(); await loadBenchmarkSnapshot(); renderMemoryState(); diff --git a/panel/panel.ts b/panel/panel.ts index 7009ddd..c292a7c 100644 --- a/panel/panel.ts +++ b/panel/panel.ts @@ -21,8 +21,10 @@ const memoryShellEl = $('#memory-shell'); const memoryStatusEl = $('#memory-status'); const memoryMetaEl = $('#memory-meta'); const memoryTargetEl = $('#memory-target') as HTMLSelectElement | null; -const orderIdInputEl = $('#order-id') as HTMLInputElement | null; -const syncOrderButtonEl = $('#btn-sync-order') as HTMLButtonElement | null; +const licenseTokenInputEl = $('#license-token') as HTMLInputElement | null; +const attachLicenseButtonEl = $('#btn-attach-license') as HTMLButtonElement | null; +const startTrialButtonEl = $('#btn-start-trial') as HTMLButtonElement | null; +const viewPlansButtonEl = $('#btn-view-plans') as HTMLButtonElement | null; const entitlementStatusEl = $('#entitlement-status'); const memoryToggleButtonEl = $('#btn-memory-toggle') as HTMLButtonElement | null; const memoryInspectButtonEl = $('#btn-memory-inspect') as HTMLButtonElement | null; @@ -62,6 +64,8 @@ const runtimeMetaOperationEl = $('#runtime-meta-operation'); const runtimeMetaStepEl = $('#runtime-meta-step'); const runtimeMetaTraceEl = $('#runtime-meta-trace'); const runtimeMetaEventsEl = $('#runtime-meta-events') as HTMLUListElement | null; +const resultShellEl = $('#result-shell'); +const processingFieldEl = $('#processing-field'); const actionButtons = Array.from(document.querySelectorAll('.primary-action, .secondary-grid button, .advanced-grid button')); type ResultView = 'readable' | 'structured'; @@ -131,6 +135,7 @@ type SelectionPreview = { title: string; url: string; hasSelection: boolean; + pageColor?: string; }; type RenderedResult = { @@ -191,9 +196,19 @@ let memorySnapshot: MemoryStatusSnapshot = { lastUpdatedAt: null, }; let entitlementSnapshot: EntitlementSnapshot = null; +let bridgeAvailable = false; +let installationSnapshot = { + state: 'idle', + label: 'Ready to install', + progress: 0, + profile: null as string | null, + action_required: null as string | null, +}; +let installationPollTimer: number | null = null; let firstRunCompleted = false; const BENCHMARK_CACHE_KEY = 'selectpilot_runtime_benchmark_v1'; const RUNTIME_META_MAX_EVENTS = 6; +const XRAY_ENABLED = new URLSearchParams(location.search).has('xray'); let runtimeMetaEventSource: EventSource | null = null; let runtimeMetaReconnectTimer: number | null = null; @@ -213,21 +228,16 @@ const runtimeMetaOverlayState: RuntimeMetaOverlayState = { let intentSuggestions: string[] = []; let selectedIntentSuggestion: string | null = null; -const FAST_INSTALL_COMMANDS = [ - 'ollama pull gemma4:e2b-it-qat', - 'ollama pull nomic-embed-text-v2-moe:latest', - 'CHROMEAI_OLLAMA_MODEL=gemma4:e2b-it-qat CHROMEAI_OLLAMA_NUM_CTX=16384 ./scripts/install-macos-local.sh', -].join('\n'); - -const QUICK_SETUP_COMMANDS = [ - 'pnpm setup:local', - 'curl http://127.0.0.1:8083/health', -].join('\n'); - function setStatus(text: string) { + if (resultShellEl?.getAttribute('aria-busy') === 'true') return; if (statusEl) statusEl.textContent = text; } +function setSilentProcessing(active: boolean) { + processingFieldEl?.classList.toggle('is-active', active); + resultShellEl?.setAttribute('aria-busy', String(active)); +} + function setStatusBar(text: string) { if (statusBar) statusBar.textContent = text; } @@ -265,13 +275,6 @@ function shorten(text: string, max = 220): string { return `${trimmed.slice(0, max - 1).trimEnd()}…`; } -function getEfficiencyScore(snapshot: BenchmarkSnapshot | null): number | null { - if (!snapshot) return null; - const avgLatency = (snapshot.extract_latency_ms + snapshot.summarize_latency_ms) / 2; - const score = Math.round(100 - avgLatency / 40); - return Math.max(0, Math.min(100, score)); -} - function formatPrivacyVerifiedAt(iso?: string): string { if (!iso) return 'Awaiting proof'; const date = new Date(iso); @@ -666,16 +669,6 @@ function getEffectiveRecommendedProfile() { return getRuntimeProfile(getEffectiveRecommendedProfileKey()); } -function getEffectiveRecommendationReason(): string { - if (!benchmarkSnapshot) return runtimeProfilesPayload.reason; - const benchmarkProfile = getRuntimeProfile(benchmarkSnapshot.recommended_profile).label; - const autoProfile = getRuntimeProfile(benchmarkSnapshot.auto_profile || runtimeProfilesPayload.recommended_profile).label; - if (benchmarkSnapshot.recommended_profile === (benchmarkSnapshot.auto_profile || runtimeProfilesPayload.recommended_profile)) { - return `Benchmark confirms the ${benchmarkProfile} profile for this workload.`; - } - return `Benchmark overrides the hardware heuristic: use ${benchmarkProfile} for this workload instead of the auto ${autoProfile} profile.`; -} - async function loadBenchmarkSnapshot() { const cached = await getJSON(BENCHMARK_CACHE_KEY); if (!cached || !cached.recommended_profile) return; @@ -908,12 +901,12 @@ function renderEntitlementStatus() { const tier = entitlementSnapshot?.tier || 'essential'; const token = entitlementSnapshot?.token; if (!token) { - entitlementStatusEl.textContent = 'No entitlement token attached yet.'; + entitlementStatusEl.textContent = 'No active access yet.'; return; } const suffix = formatEntitlementUpdatedAt(entitlementSnapshot?.cachedAt); const cacheState = getEntitlementCacheState(entitlementSnapshot); - entitlementStatusEl.textContent = `Token attached · tier ${tier} (${cacheState}) · ${suffix}`; + entitlementStatusEl.textContent = `${tier} access · ${cacheState} · ${suffix}`; } async function refreshEntitlementStatus() { @@ -927,35 +920,37 @@ async function refreshEntitlementStatus() { syncControlAvailability(); } -async function doSyncOrderToken() { - const orderId = orderIdInputEl?.value.trim() || ''; - if (!orderId) throw new Error('Order ID is required'); - - setStatus('Checking order status...'); - const orderRes = await fetch(endpoints.billingOrderStatus(orderId), { cache: 'no-store' }); - if (!orderRes.ok) { - if (orderRes.status === 404) throw new Error('No payment detected for this order ID'); - throw new Error(`Order lookup failed (${orderRes.status})`); - } - const order = await orderRes.json() as { paid?: boolean; token?: string; confirmations?: number; confirmations_required?: number }; - - if (!order.paid || !order.token) { - const confirmations = order.confirmations || 0; - const needed = order.confirmations_required || 0; - throw new Error(`No payment detected yet (${confirmations}/${needed} confirmations)`); - } - - if (entitlementSnapshot?.token && entitlementSnapshot.token === order.token) { - await request('entitlement:refresh'); - await refreshEntitlementStatus(); - setStatus('Order already synced; entitlement refreshed'); - return; - } - - await request('license:attach_token', { token: order.token }); +async function attachLicenseToken(token = licenseTokenInputEl?.value.trim() || '') { + if (!token) throw new Error('Enter your license key'); + setStatus('Activating access...'); + await request('license:attach_token', { token }); await request('entitlement:refresh'); await refreshEntitlementStatus(); - setStatus('Token attached and entitlement refreshed'); + if (licenseTokenInputEl) licenseTokenInputEl.value = ''; + setStatus('Access ready'); +} + +async function getInstallationId(): Promise { + const key = 'selectpilot_installation_id'; + const stored = await chrome.storage.local.get(key); + if (typeof stored[key] === 'string' && stored[key]) return stored[key]; + const value = crypto.randomUUID(); + await chrome.storage.local.set({ [key]: value }); + return value; +} + +async function startTrial() { + setStatus('Preparing your trial...'); + const response = await fetch(endpoints.licenseTrial, { + method: 'POST', + headers: { 'Content-Type': 'application/json' }, + body: JSON.stringify({ installation_id: await getInstallationId() }), + cache: 'no-store', + }); + if (!response.ok) throw new Error('Trial could not be started'); + const result = await response.json() as { token?: string }; + if (!result.token) throw new Error('Trial response was incomplete'); + await attachLicenseToken(result.token); } async function refreshMemoryStatus() { @@ -1001,10 +996,8 @@ function syncControlAvailability() { if (memoryInspectButtonEl) memoryInspectButtonEl.disabled = memoryActionsLocked; if (memoryExportButtonEl) memoryExportButtonEl.disabled = isBusy || !flowExportSupported || !canExport; if (memoryDeleteButtonEl) memoryDeleteButtonEl.disabled = memoryActionsLocked || memorySnapshot.entries === 0; - if (syncOrderButtonEl) { - const hasOrderId = Boolean(orderIdInputEl?.value.trim()); - syncOrderButtonEl.disabled = isBusy || !hasOrderId; - } + if (attachLicenseButtonEl) attachLicenseButtonEl.disabled = isBusy || !licenseTokenInputEl?.value.trim(); + if (startTrialButtonEl) startTrialButtonEl.disabled = isBusy || Boolean(entitlementSnapshot?.token); if (intentExecuteButtonEl) { const hasIntent = Boolean(intentInputEl?.value.trim()); intentExecuteButtonEl.disabled = isBusy || !runtimeReady || !selectionReady || !hasIntent; @@ -1031,225 +1024,94 @@ function syncPresetHelp() { if (extractHelpEl) extractHelpEl.textContent = preset.description; } -function buildMetric(label: string, value: string) { - const metric = document.createElement('div'); - metric.className = 'runtime-metric'; - const kicker = document.createElement('span'); - kicker.className = 'section-kicker'; - kicker.textContent = label; - const strong = document.createElement('strong'); - strong.textContent = value; - metric.append(kicker, strong); - return metric; -} - -function createProfileCard(profile: RuntimeProfile, recommendedKey: string, copyLabel = 'Copy command') { - const card = document.createElement('div'); - card.className = 'profile-card'; - if (profile.key === recommendedKey) card.classList.add('is-recommended'); - - const header = document.createElement('div'); - header.className = 'profile-header'; - - const title = document.createElement('div'); - title.className = 'profile-title'; - title.textContent = profile.label; - - header.append(title); - if (profile.key === recommendedKey) { - const badge = document.createElement('span'); - badge.className = 'profile-badge'; - badge.textContent = 'Recommended'; - header.append(badge); - } - - const description = document.createElement('p'); - description.className = 'selection-copy'; - description.textContent = profile.description; - - const stack = document.createElement('div'); - stack.className = 'profile-stack'; - stack.append( - buildMetric('Generation', profile.generation_model), - buildMetric('Embedding', profile.embedding_model), - buildMetric('Latency', profile.target_latency), - ); - - const intendedFor = document.createElement('p'); - intendedFor.className = 'selection-copy'; - intendedFor.textContent = profile.intended_for; - - const actions = document.createElement('div'); - actions.className = 'runtime-actions'; - const copyButton = document.createElement('button'); - copyButton.type = 'button'; - copyButton.textContent = copyLabel; - copyButton.addEventListener('click', async () => { - await navigator.clipboard.writeText(profile.command); - setStatus(`${profile.label} bootstrap command copied`); - }); - actions.append(copyButton); - - card.append(header, description, stack, intendedFor, actions); - return card; -} - function renderRuntimeState() { clearNode(runtimeStateEl); if (!runtimeStateEl) return; runtimeStateEl.classList.add('is-visible'); - const profilesGrid = document.createElement('div'); - profilesGrid.className = 'profile-grid'; - const effectiveRecommendedProfileKey = getEffectiveRecommendedProfileKey(); - for (const profile of runtimeProfilesPayload.profiles) { - profilesGrid.append(createProfileCard(profile, effectiveRecommendedProfileKey)); + if (runtimeSnapshot.ok) { + runtimeStateEl.classList.remove('is-visible'); + return; } - if (runtimeSnapshot.ok) { - const wrapper = document.createElement('div'); - wrapper.className = 'runtime-grid'; - wrapper.append( - buildMetric('Execution', 'Local-only ready'), - buildMetric('Provider', 'Ollama on-device'), - buildMetric('Ignored remote', `${runtimeSnapshot.ignoredRemoteCount} models`), + if (!bridgeAvailable) { + const helper = createCard( + 'Installation', + 'One small helper is needed', + 'It keeps selected text on this computer and prepares the right local model.' ); + const link = document.createElement('a'); + link.className = 'primary-action'; + link.href = 'https://selectpilot.app/downloads/SelectPilot-Installer.pkg'; + link.target = '_blank'; + link.rel = 'noreferrer'; + link.textContent = 'Install local helper'; + runtimeStateEl.append(helper, link); + return; + } - const copy = document.createElement('p'); - copy.className = 'runtime-copy'; - copy.textContent = `Runtime ready. ${runtimeSnapshot.activeModel} is active locally, and the selected-text path stays inside the local bridge.`; - - const benchmarkBlock = document.createElement('div'); - benchmarkBlock.className = 'benchmark-block'; - const benchmarkHeader = document.createElement('div'); - benchmarkHeader.className = 'profile-header'; - const benchmarkTitle = document.createElement('div'); - benchmarkTitle.className = 'profile-title'; - benchmarkTitle.textContent = 'Profile fit'; - benchmarkHeader.append(benchmarkTitle); - benchmarkBlock.append(benchmarkHeader); - - const benchmarkCopy = document.createElement('p'); - benchmarkCopy.className = 'selection-copy'; - if (benchmarkSnapshot) { - benchmarkCopy.textContent = getEffectiveRecommendationReason(); - } else { - benchmarkCopy.textContent = 'Run a local benchmark to confirm that the current runtime matches the smallest viable profile for this machine.'; - } - benchmarkBlock.append(benchmarkCopy); - - const benchmarkMetrics = document.createElement('div'); - benchmarkMetrics.className = 'runtime-grid'; - const efficiencyScore = getEfficiencyScore(benchmarkSnapshot); - if (benchmarkSnapshot) { - benchmarkMetrics.append( - buildMetric('Extract JSON', `${benchmarkSnapshot.extract_latency_ms} ms`), - buildMetric('Summarize', `${benchmarkSnapshot.summarize_latency_ms} ms`), - buildMetric('Efficiency', `${efficiencyScore}/100`), - buildMetric('Recommended', getEffectiveRecommendedProfile().label), - ); - } else { - benchmarkMetrics.append( - buildMetric('Extract JSON', 'Pending'), - buildMetric('Summarize', 'Pending'), - buildMetric('Efficiency', 'Pending'), - buildMetric('Recommended', getEffectiveRecommendedProfile().label), - ); - } - benchmarkBlock.append(benchmarkMetrics); - - const benchmarkActions = document.createElement('div'); - benchmarkActions.className = 'runtime-actions'; - const benchmarkButton = document.createElement('button'); - benchmarkButton.type = 'button'; - benchmarkButton.id = 'btn-run-benchmark'; - benchmarkButton.textContent = benchmarkSnapshot ? 'Run benchmark again' : 'Run benchmark'; - benchmarkActions.append(benchmarkButton); - benchmarkBlock.append(benchmarkActions); - - runtimeStateEl.append(wrapper, copy, benchmarkBlock, profilesGrid); + const installation = createCard( + 'Installation', + installationSnapshot.label, + 'The right local setup will be chosen automatically for this Mac.' + ); + const progress = document.createElement('div'); + progress.className = 'installation-progress'; + progress.setAttribute('role', 'progressbar'); + progress.setAttribute('aria-valuemin', '0'); + progress.setAttribute('aria-valuemax', '100'); + progress.setAttribute('aria-valuenow', String(installationSnapshot.progress)); + const fill = document.createElement('span'); + fill.style.width = `${installationSnapshot.progress}%`; + progress.append(fill); + installation.append(progress); + + if (installationSnapshot.state === 'installing') { + runtimeStateEl.append(installation); + scheduleInstallationPoll(); return; } - const header = document.createElement('div'); - header.className = 'output-card'; - header.innerHTML = '
    Human intervention

    Local runtime not ready

    '; + const action = document.createElement('button'); + action.className = 'primary-action'; + action.type = 'button'; + action.textContent = installationSnapshot.state === 'action_required' ? 'Try again' : 'Install Ollama and continue'; + action.addEventListener('click', () => void startLocalInstallation()); + runtimeStateEl.append(installation, action); + return; - const copy = document.createElement('p'); - copy.className = 'runtime-copy'; - copy.textContent = - 'SelectPilot runs only through local Ollama. Follow these three tiny steps and press re-check when done.'; - - const tutorialSteps = document.createElement('div'); - tutorialSteps.className = 'tutorial-steps'; - - const stepInstall = document.createElement('div'); - stepInstall.className = 'tutorial-step'; - stepInstall.innerHTML = 'Step 1Copy and run the one-command setup in Terminal.'; - - const stepLoad = document.createElement('div'); - stepLoad.className = 'tutorial-step'; - stepLoad.innerHTML = 'Step 2Open chrome://extensions, enable Developer Mode, then Load unpacked.'; - - const stepVerify = document.createElement('div'); - stepVerify.className = 'tutorial-step'; - stepVerify.innerHTML = 'Step 3Highlight any text, click Extract JSON, and confirm output appears.'; - - tutorialSteps.append(stepInstall, stepLoad, stepVerify); - - const wrapper = document.createElement('div'); - wrapper.className = 'runtime-grid'; - const recommended = getEffectiveRecommendedProfile(); - wrapper.append( - buildMetric('Recommended profile', recommended.label), - buildMetric('Generation model', recommended.generation_model), - buildMetric('Embedding model', recommended.embedding_model), - ); +} - const actions = document.createElement('div'); - actions.className = 'runtime-actions'; - - const copySetup = document.createElement('button'); - copySetup.type = 'button'; - copySetup.textContent = 'Copy one-command setup'; - copySetup.addEventListener('click', async () => { - await navigator.clipboard.writeText(QUICK_SETUP_COMMANDS); - setStatus('One-command setup copied'); - }); - actions.append(copySetup); - - const copyAdvancedSetup = document.createElement('button'); - copyAdvancedSetup.type = 'button'; - copyAdvancedSetup.textContent = 'Copy manual fallback'; - copyAdvancedSetup.addEventListener('click', async () => { - await navigator.clipboard.writeText(recommended.command || FAST_INSTALL_COMMANDS); - setStatus('Fallback setup copied'); - }); - actions.append(copyAdvancedSetup); +async function refreshInstallationStatus() { + const response = await fetch(endpoints.installationStatus, { cache: 'no-store' }); + if (!response.ok) throw new Error('installation_status_unavailable'); + installationSnapshot = await response.json(); +} - const refresh = document.createElement('button'); - refresh.type = 'button'; - refresh.textContent = 'Re-check runtime'; - refresh.addEventListener('click', () => { - void refreshRuntime(); +async function startLocalInstallation() { + const response = await fetch(endpoints.installationStart, { + method: 'POST', + headers: { 'Content-Type': 'application/json' }, + body: JSON.stringify({ consent: true }), + cache: 'no-store', }); - actions.append(refresh); - - if (runtimeSnapshot.error || runtimeSnapshot.hint) { - const note = document.createElement('p'); - note.className = 'runtime-copy'; - note.textContent = runtimeSnapshot.error || runtimeSnapshot.hint || ''; - const reason = document.createElement('p'); - reason.className = 'runtime-copy'; - reason.textContent = getEffectiveRecommendationReason(); - runtimeStateEl.append(header, copy, wrapper, tutorialSteps, actions, reason, profilesGrid, note); - return; - } + if (!response.ok) throw new Error('Installation could not start'); + installationSnapshot = await response.json(); + renderRuntimeState(); +} - const reason = document.createElement('p'); - reason.className = 'runtime-copy'; - reason.textContent = getEffectiveRecommendationReason(); - runtimeStateEl.append(header, copy, wrapper, tutorialSteps, actions, reason, profilesGrid); +function scheduleInstallationPoll() { + if (installationPollTimer !== null) return; + installationPollTimer = window.setTimeout(async () => { + installationPollTimer = null; + try { + await refreshInstallationStatus(); + if (installationSnapshot.state === 'ready') await refreshRuntime(); + else renderRuntimeState(); + } catch { + renderRuntimeState(); + } + }, 900); } function renderSelectionState() { @@ -1309,7 +1171,11 @@ async function refreshSelectionPreview() { title: preview.title || '', url: preview.url || '', hasSelection: Boolean(preview.hasSelection), + pageColor: typeof preview.pageColor === 'string' ? preview.pageColor : '', }; + if (/^(?:rgb|hsl)a?\([^)]*\)$|^#[0-9a-f]{6}$/i.test(selectionPreview.pageColor || '')) { + document.documentElement.style.setProperty('--context-tint', selectionPreview.pageColor || 'transparent'); + } setSelectionContext({ selectionOrigin: selectionPreview.hasSelection ? 'selection' : 'page_context', contentLength: selectionPreview.hasSelection ? selectionPreview.selection.length : selectionPreview.pageText.length, @@ -1338,6 +1204,12 @@ async function refreshRuntime() { }; } const health = await fetchHealth(); + bridgeAvailable = true; + try { + await refreshInstallationStatus(); + } catch { + installationSnapshot = { state: 'idle', label: 'Ready to install', progress: 0, profile: null, action_required: null }; + } privacyProofSnapshot = await fetchPrivacyProof(); runtimeSnapshot = { ok: Boolean(health?.ok), @@ -1378,6 +1250,7 @@ async function refreshRuntime() { await reconcileBenchmarkSnapshot(); renderRuntimeState(); } catch (e: any) { + bridgeAvailable = false; runtimeProfilesPayload = { profiles: RUNTIME_PROFILES, recommended_profile: 'fast', @@ -1621,17 +1494,20 @@ async function doMemoryDelete() { } function bindActions() { - const wrap = (fn: () => Promise) => async () => { + const wrap = (fn: () => Promise, mode: 'silent' | 'explicit' = 'explicit') => async () => { isBusy = true; + if (mode === 'silent') setSilentProcessing(true); syncControlAvailability(); try { await fn(); } catch (e: any) { + setSilentProcessing(false); const text = formatPanelError(e); setStatus(text); setStatusBar(text); } finally { isBusy = false; + setSilentProcessing(false); syncControlAvailability(); void Promise.all([refreshSelectionPreview(), refreshMemoryStatus(), refreshEntitlementStatus()]); } @@ -1640,20 +1516,20 @@ function bindActions() { refreshButtonEl?.addEventListener('click', () => { void Promise.all([refreshRuntime(), refreshSelectionPreview(), refreshMemoryStatus(), refreshEntitlementStatus()]); }); - $('#btn-extract')?.addEventListener('click', wrap(() => doExtract())); + $('#btn-extract')?.addEventListener('click', wrap(() => doExtract(), 'silent')); selectionCardEl?.addEventListener('click', (event) => { const target = event.target as HTMLElement | null; - if (target?.id === 'btn-first-run-example') void wrap(() => doFirstRunExample())(); + if (target?.id === 'btn-first-run-example') void wrap(() => doFirstRunExample(), 'silent')(); }); - $('#btn-summarize')?.addEventListener('click', wrap(() => doSummarize())); - $('#btn-rewrite')?.addEventListener('click', wrap(() => doRewrite())); - $('#btn-actions')?.addEventListener('click', wrap(() => doActions())); - $('#btn-ask')?.addEventListener('click', wrap(() => doAsk())); + $('#btn-summarize')?.addEventListener('click', wrap(() => doSummarize(), 'silent')); + $('#btn-rewrite')?.addEventListener('click', wrap(() => doRewrite(), 'silent')); + $('#btn-actions')?.addEventListener('click', wrap(() => doActions(), 'silent')); + $('#btn-ask')?.addEventListener('click', wrap(() => doAsk(), 'silent')); memoryToggleButtonEl?.addEventListener('click', wrap(() => doMemoryToggle())); memoryInspectButtonEl?.addEventListener('click', wrap(() => doMemoryInspect())); memoryExportButtonEl?.addEventListener('click', wrap(() => doMemoryExport())); memoryDeleteButtonEl?.addEventListener('click', wrap(() => doMemoryDelete())); - intentExecuteButtonEl?.addEventListener('click', wrap(() => doExecuteIntent())); + intentExecuteButtonEl?.addEventListener('click', wrap(() => doExecuteIntent(), 'silent')); intentClearButtonEl?.addEventListener('click', () => clearIntentInput()); intentInputEl?.addEventListener('input', () => { selectedIntentSuggestion = null; @@ -1662,10 +1538,14 @@ function bindActions() { intentInputEl?.addEventListener('keydown', (event) => { if (event.key !== 'Enter') return; event.preventDefault(); - void wrap(() => doExecuteIntent())(); + void wrap(() => doExecuteIntent(), 'silent')(); }); - syncOrderButtonEl?.addEventListener('click', wrap(() => doSyncOrderToken())); - orderIdInputEl?.addEventListener('input', () => syncControlAvailability()); + attachLicenseButtonEl?.addEventListener('click', wrap(() => attachLicenseToken())); + startTrialButtonEl?.addEventListener('click', wrap(() => startTrial())); + viewPlansButtonEl?.addEventListener('click', wrap(async () => { + await chrome.tabs.create({ url: 'https://selectpilot.app/pricing', active: true }); + })); + licenseTokenInputEl?.addEventListener('input', () => syncControlAvailability()); extractPresetEl?.addEventListener('change', () => syncPresetHelp()); runtimeStateEl?.addEventListener('click', (event) => { const target = event.target as HTMLElement | null; @@ -1694,6 +1574,14 @@ populatePresetOptions(); bindActions(); async function initialize() { + const storedSettings = await chrome.storage.sync.get('selectpilot_settings'); + const settings = (storedSettings.selectpilot_settings || {}) as Record; + const textSize = typeof settings.textSize === 'string' ? settings.textSize : 'standard'; + document.body.dataset.controlSide = settings.controlSide === 'left' ? 'left' : 'right'; + document.body.dataset.textSize = ['large', 'larger'].includes(textSize) ? textSize : 'standard'; + document.body.dataset.highContrast = settings.highContrast ? 'true' : 'false'; + document.body.dataset.reduceMotion = settings.reduceMotion ? 'true' : 'false'; + currentResultView = settings.resultView === 'structured' ? 'structured' : 'readable'; const topologyValidation = validateTopologyMap(); const requiredTopologyComponents = [ 'panel_header', @@ -1719,8 +1607,10 @@ async function initialize() { setVisiblePanels(['selection_surface', 'runtime_surface', 'report_surface']); firstRunCompleted = Boolean(await getJSON(FIRST_RUN_COMPLETED_KEY)); - renderRuntimeMetaOverlay(); - void connectRuntimeMetaStream(); + if (XRAY_ENABLED) { + renderRuntimeMetaOverlay(); + void connectRuntimeMetaStream(); + } refreshIntentSuggestions(); await loadBenchmarkSnapshot(); renderMemoryState(); diff --git a/popup/popup.js b/popup/popup.js index 24ac505..3ea4371 100644 --- a/popup/popup.js +++ b/popup/popup.js @@ -2,8 +2,18 @@ // spec_ref: "frontend_state_contract" import { log } from '../utils/logger.js'; const btn = document.getElementById('open-panel'); +async function injectIntoActiveTab() { + const [tab] = await chrome.tabs.query({ active: true, currentWindow: true }); + if (!tab?.id) + throw new Error('No active tab is available'); + await chrome.scripting.executeScript({ + target: { tabId: tab.id }, + files: ['content/content-script.bundle.js'], + }); +} btn?.addEventListener('click', async () => { try { + await injectIntoActiveTab(); await chrome.sidePanel.open({ windowId: chrome.windows.WINDOW_ID_CURRENT }); await chrome.sidePanel.setOptions({ path: 'panel/panel.html', enabled: true }); window.close(); diff --git a/popup/popup.ts b/popup/popup.ts index a9dcad4..aec88d1 100644 --- a/popup/popup.ts +++ b/popup/popup.ts @@ -4,8 +4,18 @@ import { log } from '../utils/logger.js'; const btn = document.getElementById('open-panel'); +async function injectIntoActiveTab(): Promise { + const [tab] = await chrome.tabs.query({ active: true, currentWindow: true }); + if (!tab?.id) throw new Error('No active tab is available'); + await chrome.scripting.executeScript({ + target: { tabId: tab.id }, + files: ['content/content-script.bundle.js'], + }); +} + btn?.addEventListener('click', async () => { try { + await injectIntoActiveTab(); await chrome.sidePanel.open({ windowId: chrome.windows.WINDOW_ID_CURRENT }); await chrome.sidePanel.setOptions({ path: 'panel/panel.html', enabled: true }); window.close(); diff --git a/pricing/entitlement-public-keys.json b/pricing/entitlement-public-keys.json new file mode 100644 index 0000000..0ff0fca --- /dev/null +++ b/pricing/entitlement-public-keys.json @@ -0,0 +1,11 @@ +{ + "schema_version": 1, + "keys": [ + { + "kid": "selectpilot-prod-2026-01", + "alg": "Ed25519", + "public_key_hex": "e9c1429a51ccf7d5eac11c4d22139a85414231d3ad64578e1fdb7f004ed74bd7", + "status": "active" + } + ] +} diff --git a/scripts/build-site-config.mjs b/scripts/build-site-config.mjs new file mode 100644 index 0000000..188e4d4 --- /dev/null +++ b/scripts/build-site-config.mjs @@ -0,0 +1,24 @@ +import { mkdir, readFile, writeFile } from 'node:fs/promises'; +import path from 'node:path'; + +const required = ['PADDLE_CLIENT_TOKEN', 'PADDLE_PRICE_ESSENTIAL', 'PADDLE_PRICE_PLUS', 'PADDLE_PRICE_PRO', 'ENTITLEMENT_AUTHORITY_URL', 'CHROME_WEB_STORE_URL']; +for (const name of required) if (!process.env[name]?.trim()) throw new Error(`Missing ${name}`); +for (const name of ['PADDLE_PRICE_ESSENTIAL', 'PADDLE_PRICE_PLUS', 'PADDLE_PRICE_PRO']) { + if (!/^pri_[A-Za-z0-9]+$/.test(process.env[name])) throw new Error(`${name} must be a Paddle price ID`); +} +for (const name of ['ENTITLEMENT_AUTHORITY_URL', 'CHROME_WEB_STORE_URL']) { + const url = new URL(process.env[name]); + if (url.protocol !== 'https:') throw new Error(`${name} must use HTTPS`); +} +const config = `window.SELECTPILOT_CONFIG = ${JSON.stringify({ + paddleEnvironment: process.env.PADDLE_ENVIRONMENT === 'sandbox' ? 'sandbox' : 'production', + paddleClientToken: process.env.PADDLE_CLIENT_TOKEN, + prices: { essential: process.env.PADDLE_PRICE_ESSENTIAL, plus: process.env.PADDLE_PRICE_PLUS, pro: process.env.PADDLE_PRICE_PRO }, + entitlementAuthorityUrl: process.env.ENTITLEMENT_AUTHORITY_URL, +})};\n`; +await mkdir('site-dist', { recursive: true }); +for (const file of ['index.html', 'pricing.html', 'checkout.html', 'checkout.js', 'privacy.html', 'support.html', 'styles.css']) { + const source = await readFile(path.join('site', file), 'utf8'); + await writeFile(path.join('site-dist', file), source.replaceAll('INSTALL_URL', process.env.CHROME_WEB_STORE_URL)); +} +await writeFile('site-dist/config.js', config, { mode: 0o644 }); diff --git a/scripts/lint-manifest.mjs b/scripts/lint-manifest.mjs index 4fe5b82..a840448 100644 --- a/scripts/lint-manifest.mjs +++ b/scripts/lint-manifest.mjs @@ -46,6 +46,14 @@ if (Array.isArray(manifest.content_scripts)) { }); } +if (manifest.permissions?.includes('tabs')) { + errors.push('permissions: tabs är inte tillåten; använd activeTab för det användarinitierade flödet.'); +} + +if (manifest.content_scripts?.some((entry) => entry?.matches?.includes(''))) { + errors.push('content_scripts: permanent -injektion är inte tillåten; använd activeTab + scripting.'); +} + if (Array.isArray(manifest.web_accessible_resources)) { manifest.web_accessible_resources.forEach((entry, index) => { validateMatches(entry?.matches, `web_accessible_resources[${index}]`); @@ -60,4 +68,4 @@ if (errors.length > 0) { process.exit(1); } -console.log('Manifest lint OK'); \ No newline at end of file +console.log('Manifest lint OK'); diff --git a/scripts/package-chrome-store.mjs b/scripts/package-chrome-store.mjs index 4341837..626e210 100644 --- a/scripts/package-chrome-store.mjs +++ b/scripts/package-chrome-store.mjs @@ -11,7 +11,7 @@ import { validateStoreAssets } from './validate-store-assets.mjs'; const projectRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..'); const fixedDate = new Date('2026-01-01T00:00:00Z'); -const runtimeRoots = ['agent', 'api', 'background', 'content', 'licensing', 'panel', 'popup', 'shared', 'utils']; +const runtimeRoots = ['agent', 'api', 'background', 'content', 'licensing', 'options', 'panel', 'popup', 'shared', 'utils']; const runtimeAssets = [ 'assets/icon16.png', 'assets/icon32.png', @@ -20,6 +20,7 @@ const runtimeAssets = [ 'assets/icon256.png', 'assets/icon512.png', 'pricing/tier-feature-map.json', + 'pricing/entitlement-public-keys.json', ]; const allowedExtensions = new Set(['.css', '.html', '.js', '.json', '.png', '.svg']); const forbiddenPath = /(^|\/)(?:\.env|tests?|reports?|logs?|node_modules|test-results|playwright-report)(\/|$)|\.(?:map|log|pem|key)$/i; @@ -27,6 +28,11 @@ const secretPatterns = [ /-----BEGIN (?:RSA |EC |OPENSSH )?PRIVATE KEY-----/, /(?:api|secret|private)[_-]?key\s*[:=]\s*["'][^"']{12,}["']/i, ]; +const remoteCodePatterns = [ + /]+src=["']https?:\/\//i, + /\.src\s*=\s*["']https?:\/\/[^"']+\.js(?:[?"'])/i, + /import\s*\(\s*["']https?:\/\//i, +]; async function walk(directory, base = directory) { const entries = await readdir(directory, { withFileTypes: true }); @@ -56,6 +62,21 @@ export async function assertReleaseSafe(files, root = projectRoot) { if (entitlement.includes('__SELECTPILOT_ENTITLEMENT_PUBLIC_KEY_HEX__') || entitlement.includes('__SELECTPILOT_ENTITLEMENT_KEY_ID__')) { throw new Error('Store package blocked: production entitlement signature verification is not configured (SOD-837).'); } + const keyring = JSON.parse(await readFile(path.join(root, 'pricing/entitlement-public-keys.json'), 'utf8')); + if ( + keyring.schema_version !== 1 + || !Array.isArray(keyring.keys) + || !keyring.keys.some((key) => key.status === 'active') + || keyring.keys.some((key) => ( + typeof key.kid !== 'string' + || !key.kid + || key.alg !== 'Ed25519' + || !/^[0-9a-f]{64}$/i.test(key.public_key_hex || '') + || !['active', 'retiring'].includes(key.status) + )) + ) { + throw new Error('Store package blocked: production entitlement keyring is invalid.'); + } if (files.some((file) => file.startsWith('billing/') || file === 'pricing/paddle-products.json')) { throw new Error('Store package blocked: inactive remote checkout code or product placeholders entered runtime inventory.'); } @@ -69,6 +90,9 @@ export async function assertReleaseSafe(files, root = projectRoot) { for (const pattern of secretPatterns) { if (pattern.test(content)) throw new Error(`Potential secret in release file: ${relative}`); } + for (const pattern of remoteCodePatterns) { + if (pattern.test(content)) throw new Error(`Remote hosted code in release file: ${relative}`); + } } } } diff --git a/scripts/package-macos-helper.sh b/scripts/package-macos-helper.sh new file mode 100644 index 0000000..ed9e713 --- /dev/null +++ b/scripts/package-macos-helper.sh @@ -0,0 +1,37 @@ +#!/bin/sh +set -eu + +ROOT="$(CDPATH= cd -- "$(dirname "$0")/.." && pwd)" +STAGE="$ROOT/dist/macos-helper/payload" +OUTPUT="$ROOT/dist/macos-helper/SelectPilot-Installer-unsigned.pkg" +SIGNED="$ROOT/dist/macos-helper/SelectPilot-Installer.pkg" +INSTALL_ROOT="$STAGE/Library/Application Support/SelectPilot" + +rm -rf "$ROOT/dist/macos-helper" +mkdir -p "$INSTALL_ROOT" "$INSTALL_ROOT/launchd" +for source_dir in server presets runtime; do + find "$ROOT/$source_dir" -type f \( -name '*.py' -o -name '*.json' \) | while IFS= read -r source; do + relative="${source#"$ROOT/"}" + destination="$INSTALL_ROOT/$relative" + mkdir -p "$(dirname "$destination")" + cp "$source" "$destination" + done +done +cp "$ROOT/installer/macos/com.selectpilot.bridge.plist.template" "$INSTALL_ROOT/launchd/" +chmod 755 "$ROOT/installer/macos/scripts/postinstall" + +pkgbuild \ + --root "$STAGE" \ + --scripts "$ROOT/installer/macos/scripts" \ + --identifier com.selectpilot.helper \ + --version "$(node -p "require('$ROOT/package.json').version")" \ + --install-location / \ + "$OUTPUT" + +if [ -n "${SELECTPILOT_INSTALLER_SIGN_IDENTITY:-}" ]; then + productsign --sign "$SELECTPILOT_INSTALLER_SIGN_IDENTITY" "$OUTPUT" "$SIGNED" + pkgutil --check-signature "$SIGNED" + printf '%s\n' "$SIGNED" +else + printf '%s\n' "$OUTPUT" +fi diff --git a/scripts/verify-entitlement-key.mjs b/scripts/verify-entitlement-key.mjs new file mode 100644 index 0000000..a848f9c --- /dev/null +++ b/scripts/verify-entitlement-key.mjs @@ -0,0 +1,16 @@ +import { createPrivateKey, createPublicKey, sign, verify } from 'node:crypto'; +import { readFile } from 'node:fs/promises'; + +const privateKeyFile = process.env.SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE; +if (!privateKeyFile) throw new Error('SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE is required'); +const keyring = JSON.parse(await readFile('pricing/entitlement-public-keys.json', 'utf8')); +const active = keyring.keys.filter((key) => key.status === 'active'); +if (active.length !== 1) throw new Error('Exactly one active entitlement key is required'); +const privateKey = createPrivateKey(await readFile(privateKeyFile)); +const publicDer = createPublicKey(privateKey).export({ type: 'spki', format: 'der' }); +const publicHex = publicDer.subarray(-32).toString('hex'); +if (publicHex !== active[0].public_key_hex) throw new Error('Private key does not match the pinned public key'); +const payload = Buffer.from(JSON.stringify({ token: 'verification-only', tier: 'essential', features: [], issuedAt: 0, expiresAt: null })); +const signature = sign(null, payload, privateKey); +if (!verify(null, payload, createPublicKey(privateKey), signature)) throw new Error('Ed25519 verification failed'); +console.log(`Entitlement signing key verified: ${active[0].kid}`); diff --git a/server/installation_manager.py b/server/installation_manager.py new file mode 100644 index 0000000..a0f2810 --- /dev/null +++ b/server/installation_manager.py @@ -0,0 +1,155 @@ +"""Calm, explicit local runtime provisioning for macOS.""" +from __future__ import annotations + +import json +import os +import platform +import shutil +import subprocess +import tempfile +import threading +import time +from pathlib import Path +from urllib.request import Request, urlopen + +from runtime_profiles import get_runtime_profile, recommend_runtime_profile + +OLLAMA_DOWNLOAD_URL = "https://ollama.com/download/Ollama-darwin.zip" +OLLAMA_TEAM_ID = "3MU9H2V9Y9" +OLLAMA_IDENTIFIER = "com.electron.ollama" + + +class InstallationManager: + def __init__(self) -> None: + self._lock = threading.Lock() + self._state = { + "state": "idle", + "label": "Ready to install", + "progress": 0, + "profile": None, + "action_required": None, + } + + def status(self) -> dict: + with self._lock: + return dict(self._state) + + def _update(self, **values: object) -> None: + with self._lock: + self._state.update(values) + + def start(self, consent: bool) -> dict: + if consent is not True: + raise ValueError("installation_consent_required") + if platform.system() != "Darwin": + raise RuntimeError("macos_required") + with self._lock: + if self._state["state"] == "installing": + return dict(self._state) + self._state.update(state="installing", label="Preparing", progress=4, action_required=None) + threading.Thread(target=self._run, name="selectpilot-installation", daemon=True).start() + return self.status() + + def _ollama_cli(self) -> Path | None: + candidates = [ + Path("/Applications/Ollama.app/Contents/Resources/ollama"), + Path.home() / "Applications/Ollama.app/Contents/Resources/ollama", + ] + executable = shutil.which("ollama") + if executable: + candidates.insert(0, Path(executable)) + return next((path for path in candidates if path.is_file() and os.access(path, os.X_OK)), None) + + def _install_ollama(self) -> Path: + existing = self._ollama_cli() + if existing: + return existing + self._update(label="Installing", progress=12) + with tempfile.TemporaryDirectory(prefix="selectpilot-ollama-") as temporary: + root = Path(temporary) + archive = root / "Ollama-darwin.zip" + with urlopen(OLLAMA_DOWNLOAD_URL, timeout=60) as response, archive.open("wb") as output: + total = int(response.headers.get("Content-Length") or 0) + received = 0 + while True: + chunk = response.read(1024 * 1024) + if not chunk: + break + output.write(chunk) + received += len(chunk) + if total: + self._update(progress=min(38, 12 + round(26 * received / total))) + subprocess.run(["/usr/bin/ditto", "-x", "-k", str(archive), str(root)], check=True) + application = root / "Ollama.app" + requirement = ( + f'anchor apple generic and certificate leaf[subject.OU] = "{OLLAMA_TEAM_ID}" ' + f'and identifier "{OLLAMA_IDENTIFIER}"' + ) + subprocess.run( + ["/usr/bin/codesign", "--verify", "--deep", "--strict", f"-R={requirement}", str(application)], + check=True, + stdout=subprocess.DEVNULL, + stderr=subprocess.DEVNULL, + ) + destination_root = Path.home() / "Applications" + destination_root.mkdir(mode=0o700, parents=True, exist_ok=True) + destination = destination_root / "Ollama.app" + if destination.exists(): + shutil.rmtree(destination) + shutil.move(str(application), str(destination)) + return destination / "Contents/Resources/ollama" + + def _wait_for_ollama(self) -> None: + for _ in range(60): + try: + with urlopen("http://127.0.0.1:11434/api/version", timeout=1) as response: + if response.status == 200: + return + except Exception: + time.sleep(1) + raise RuntimeError("ollama_start_timeout") + + def _pull_model(self, model: str, start: int, end: int) -> None: + request = Request( + "http://127.0.0.1:11434/api/pull", + data=json.dumps({"model": model, "stream": True}).encode("utf-8"), + headers={"Content-Type": "application/json"}, + method="POST", + ) + with urlopen(request, timeout=3600) as response: + for line in response: + event = json.loads(line) + total = event.get("total") + completed = event.get("completed") + if isinstance(total, int) and total > 0 and isinstance(completed, int): + self._update(progress=min(end, start + round((end - start) * completed / total))) + if event.get("error"): + raise RuntimeError("model_download_failed") + + def _run(self) -> None: + try: + recommendation = recommend_runtime_profile() + profile = get_runtime_profile(recommendation["recommended_profile"]) + self._update(profile=profile.key) + ollama = self._install_ollama() + self._update(label="Starting local processing", progress=42) + subprocess.run(["/usr/bin/open", str(ollama.parents[2]), "--args", "hidden"], check=True) + self._wait_for_ollama() + self._update(label="Optimizing for this Mac", progress=48) + self._pull_model(profile.generation_model, 48, 82) + self._pull_model(profile.embedding_model, 82, 94) + self._update(label="Final checks", progress=96) + time.sleep(0.2) + self._update(state="ready", label="Go", progress=100) + except Exception as error: + code = str(error) if str(error) in { + "ollama_start_timeout", "model_download_failed", "macos_required" + } else "installation_failed" + self._update( + state="action_required", + label="A little more time is needed", + action_required=code, + ) + + +INSTALLATION = InstallationManager() diff --git a/server/nano_server.py b/server/nano_server.py index 1cb2de0..6969784 100644 --- a/server/nano_server.py +++ b/server/nano_server.py @@ -25,6 +25,7 @@ from ollama_client import OllamaClient, OllamaError from extraction_presets import get_extraction_preset from runtime_profiles import build_bootstrap_commands, list_runtime_profiles, recommend_runtime_profile +from installation_manager import INSTALLATION DEFAULT_PORT = 8083 LOCAL_HOSTS = {"127.0.0.1", "localhost", "::1"} @@ -44,6 +45,10 @@ "/agent", "/embed", "/license/verify", + "/license/trial", + "/license/claim", + "/installation/status", + "/installation/start", ] @@ -92,12 +97,30 @@ class OperationContract: template="license_verify.v1", allowed_fields=("token",), ), + "/license/trial": OperationContract( + name="license_trial", + endpoint="/license/trial", + template="license_trial.v1", + allowed_fields=("installation_id",), + ), + "/license/claim": OperationContract( + name="license_claim", + endpoint="/license/claim", + template="license_claim.v1", + allowed_fields=("claim_id",), + ), "/benchmark": OperationContract( name="benchmark", endpoint="/benchmark", template="benchmark.v1", allowed_fields=(), ), + "/installation/start": OperationContract( + name="installation_start", + endpoint="/installation/start", + template="installation_start.v1", + allowed_fields=("consent",), + ), } @@ -1261,26 +1284,29 @@ def benchmark_runtime() -> dict: } -def license_verify(payload: dict) -> dict: - token = str(payload.get("token") or "").strip() - if not token: - raise ValidationError("missing_token", "Entitlement token is required") - - verifier_url = os.environ.get( - "SELECTPILOT_BILLING_VERIFY_URL", - "http://127.0.0.1:8090/license/verify", - ) - parsed = urlparse(verifier_url) - if parsed.scheme != "http" or parsed.hostname not in LOCAL_HOSTS or parsed.path != "/license/verify": +def _entitlement_authority_url(path: str) -> str: + base_url = os.environ.get( + "SELECTPILOT_ENTITLEMENT_AUTHORITY_URL", + "http://127.0.0.1:8090", + ).rstrip("/") + parsed = urlparse(base_url) + is_local_http = parsed.scheme == "http" and parsed.hostname in LOCAL_HOSTS + is_remote_https = parsed.scheme == "https" and bool(parsed.hostname) + if not (is_local_http or is_remote_https) or parsed.path not in {"", "/"} or parsed.query or parsed.fragment or parsed.username: raise ValidationError( "invalid_entitlement_verifier", - "Entitlement verifier must be the explicit local HTTP license endpoint", + "Entitlement authority must use HTTPS or an explicit local loopback endpoint", status=503, ) + return f"{base_url}{path}" + + +def _entitlement_request(path: str, payload: dict) -> dict: + verifier_url = _entitlement_authority_url(path) request = Request( verifier_url, - data=json.dumps({"token": token}, separators=(",", ":")).encode("utf-8"), + data=json.dumps(payload, separators=(",", ":")).encode("utf-8"), headers={"Content-Type": "application/json"}, method="POST", ) @@ -1290,9 +1316,18 @@ def license_verify(payload: dict) -> dict: except HTTPError as exc: if exc.code == 401: raise ValidationError("invalid_entitlement", "Entitlement token is invalid", status=401) from exc - raise ValidationError("entitlement_verifier_error", "Local entitlement verifier rejected the request", status=503) from exc + raise ValidationError("entitlement_verifier_error", "Entitlement authority rejected the request", status=503) from exc except (URLError, TimeoutError, json.JSONDecodeError) as exc: - raise ValidationError("entitlement_verifier_unavailable", "Local entitlement verifier is unavailable", status=503) from exc + raise ValidationError("entitlement_verifier_unavailable", "Entitlement authority is unavailable", status=503) from exc + if not isinstance(result, dict): + raise ValidationError("invalid_entitlement_response", "Entitlement authority returned an invalid contract", status=503) + return result + +def license_verify(payload: dict) -> dict: + token = str(payload.get("token") or "").strip() + if not token: + raise ValidationError("missing_token", "Entitlement token is required") + result = _entitlement_request("/v1/entitlements/verify", {"token": token}) entitlement = result.get("entitlement") if isinstance(result, dict) else None issued_at = entitlement.get("issuedAt") if isinstance(entitlement, dict) else None @@ -1302,6 +1337,7 @@ def license_verify(payload: dict) -> dict: or not isinstance(entitlement, dict) or entitlement.get("token") != token or entitlement.get("tier") not in {"essential", "plus", "pro"} + or not isinstance(entitlement.get("features"), list) or not isinstance(issued_at, int) or isinstance(issued_at, bool) or (expires_at is not None and (not isinstance(expires_at, int) or isinstance(expires_at, bool))) @@ -1311,10 +1347,24 @@ def license_verify(payload: dict) -> dict: or not isinstance(result.get("signature"), str) or not result.get("signature") ): - raise ValidationError("invalid_entitlement_response", "Local entitlement verifier returned an invalid contract", status=503) + raise ValidationError("invalid_entitlement_response", "Entitlement authority returned an invalid contract", status=503) return result +def license_trial(payload: dict) -> dict: + installation_id = str(payload.get("installation_id") or "").strip() + if not installation_id: + raise ValidationError("missing_installation_id", "Installation identity is required") + return _entitlement_request("/v1/trials/start", {"installation_id": installation_id}) + + +def license_claim(payload: dict) -> dict: + claim_id = str(payload.get("claim_id") or "").strip() + if not claim_id: + raise ValidationError("missing_claim_id", "Purchase claim is required") + return _entitlement_request("/v1/claims/redeem", {"claim_id": claim_id}) + + class Handler(BaseHTTPRequestHandler): server_version = "ChromeAINano/1.0" @@ -1373,6 +1423,9 @@ def do_GET(self): if path == '/profiles': self._write_json(200, runtime_profiles()) return + if path == '/installation/status': + self._write_json(200, INSTALLATION.status()) + return if path == '/runtime-meta/health': self._write_json(200, { "ok": True, @@ -1622,8 +1675,19 @@ def _execute_summarize(): runtime_model_id = str(resp.get("model") or "") if isinstance(resp, dict) else None elif path == '/license/verify': resp = license_verify(payload) + elif path == '/license/trial': + resp = license_trial(payload) + elif path == '/license/claim': + resp = license_claim(payload) elif path == '/benchmark': resp = benchmark_runtime() + elif path == '/installation/start': + try: + resp = INSTALLATION.start(payload.get("consent") is True) + except ValueError as exc: + raise ValidationError("installation_consent_required", "Installation requires your approval") from exc + except RuntimeError as exc: + raise ValidationError("installation_unavailable", "Installation is not available on this device", status=422) from exc emit_runtime_meta( event_type="STEP_COMPLETED", @@ -1830,7 +1894,7 @@ def main(): expected = args.binary_hash or os.environ.get('CHROMEAI_BINARY_HASH') verify_binary(binary_path, expected) - auto_pull = os.environ.get('CHROMEAI_AUTO_PULL_MODELS', '1').strip().lower() not in {'0', 'false', 'no'} + auto_pull = os.environ.get('CHROMEAI_AUTO_PULL_MODELS', '0').strip().lower() not in {'0', 'false', 'no'} if auto_pull: ensure_runtime_models() diff --git a/services/entitlement-authority/Dockerfile b/services/entitlement-authority/Dockerfile new file mode 100644 index 0000000..d6f18fc --- /dev/null +++ b/services/entitlement-authority/Dockerfile @@ -0,0 +1,7 @@ +FROM node:22-alpine +WORKDIR /app +COPY services/entitlement-authority/server.mjs ./server.mjs +USER node +ENV HOST=0.0.0.0 PORT=8090 +EXPOSE 8090 +CMD ["node", "server.mjs"] diff --git a/services/entitlement-authority/compose.yaml b/services/entitlement-authority/compose.yaml new file mode 100644 index 0000000..2629985 --- /dev/null +++ b/services/entitlement-authority/compose.yaml @@ -0,0 +1,39 @@ +services: + authority: + build: + context: ../.. + dockerfile: services/entitlement-authority/Dockerfile + restart: unless-stopped + user: "${SELECTPILOT_UID:-1000}:${SELECTPILOT_GID:-1000}" + read_only: true + ports: + - "127.0.0.1:8090:8090" + environment: + HOST: 0.0.0.0 + PORT: "8090" + SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE: /run/secrets/entitlement_private_key + SELECTPILOT_ENTITLEMENT_KEY_ID: selectpilot-prod-2026-01 + SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE: /run/secrets/paddle_webhook_secret + SELECTPILOT_PADDLE_PRICE_MAP_FILE: /run/secrets/paddle_price_map + SELECTPILOT_ENTITLEMENT_STATE_FILE: /var/lib/selectpilot/entitlements.json + SELECTPILOT_WEB_ORIGIN: https://selectpilot.app + secrets: + - entitlement_private_key + - paddle_webhook_secret + - paddle_price_map + volumes: + - ${SELECTPILOT_STATE_DIR}:/var/lib/selectpilot + tmpfs: + - /tmp:size=16m,mode=1777 + security_opt: + - no-new-privileges:true + cap_drop: + - ALL + +secrets: + entitlement_private_key: + file: ${SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE} + paddle_webhook_secret: + file: ${SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE} + paddle_price_map: + file: ${SELECTPILOT_PADDLE_PRICE_MAP_FILE} diff --git a/services/entitlement-authority/server.mjs b/services/entitlement-authority/server.mjs new file mode 100644 index 0000000..4a0e888 --- /dev/null +++ b/services/entitlement-authority/server.mjs @@ -0,0 +1,285 @@ +import { createHash, createHmac, randomBytes, sign, timingSafeEqual } from 'node:crypto'; +import { createServer } from 'node:http'; +import { readFile, rename, writeFile } from 'node:fs/promises'; +import { dirname, resolve } from 'node:path'; +import { mkdir } from 'node:fs/promises'; + +const TIERS = new Set(['essential', 'plus', 'pro']); +const MAX_BODY_BYTES = 256 * 1024; +const PADDLE_TOLERANCE_SECONDS = 300; + +function required(name, env = process.env) { + const value = env[name]?.trim(); + if (!value) throw new Error(`Missing required environment variable: ${name}`); + return value; +} + +function canonicalize(entitlement) { + return JSON.stringify({ + token: entitlement.token, + tier: entitlement.tier, + features: Array.isArray(entitlement.features) ? entitlement.features : [], + issuedAt: entitlement.issuedAt, + expiresAt: entitlement.expiresAt ?? null, + }); +} + +function tokenHash(token) { + return createHash('sha256').update(token, 'utf8').digest('hex'); +} + +function opaque(prefix) { + return `${prefix}_${randomBytes(24).toString('base64url')}`; +} + +function parsePriceMap(value) { + const parsed = JSON.parse(value); + for (const [priceId, entry] of Object.entries(parsed)) { + if ((priceId !== '__trial__' && !priceId.startsWith('pri_')) || !TIERS.has(entry?.tier) || !Array.isArray(entry.features)) { + throw new Error('SELECTPILOT_PADDLE_PRICE_MAP_JSON has an invalid entry'); + } + } + return parsed; +} + +export function verifyPaddleSignature(rawBody, header, secret, nowSeconds = Math.floor(Date.now() / 1000)) { + const parts = Object.fromEntries(String(header || '').split(';').map((part) => part.split('=', 2))); + const timestamp = Number(parts.ts); + const supplied = parts.h1; + if (!Number.isInteger(timestamp) || !/^[0-9a-f]{64}$/i.test(supplied || '')) return false; + if (Math.abs(nowSeconds - timestamp) > PADDLE_TOLERANCE_SECONDS) return false; + const expected = createHmac('sha256', secret).update(`${timestamp}:${rawBody}`).digest('hex'); + return timingSafeEqual(Buffer.from(expected, 'hex'), Buffer.from(supplied, 'hex')); +} + +async function readBody(req) { + const chunks = []; + let size = 0; + for await (const chunk of req) { + size += chunk.length; + if (size > MAX_BODY_BYTES) throw Object.assign(new Error('request_too_large'), { status: 413 }); + chunks.push(chunk); + } + return Buffer.concat(chunks).toString('utf8'); +} + +async function loadState(path) { + try { + const parsed = JSON.parse(await readFile(path, 'utf8')); + return { schema_version: 1, tokens: {}, claims: {}, trials: {}, events: {}, subjects: {}, ...parsed }; + } catch (error) { + if (error.code === 'ENOENT') return { schema_version: 1, tokens: {}, claims: {}, trials: {}, events: {}, subjects: {} }; + throw error; + } +} + +async function saveState(path, state) { + await mkdir(dirname(path), { recursive: true }); + const temporary = `${path}.${process.pid}.tmp`; + await writeFile(temporary, `${JSON.stringify(state)}\n`, { mode: 0o600 }); + await rename(temporary, path); +} + +function send(res, status, body) { + const payload = JSON.stringify(body); + res.writeHead(status, { + 'content-type': 'application/json; charset=utf-8', + 'cache-control': 'no-store', + 'content-length': Buffer.byteLength(payload), + 'x-content-type-options': 'nosniff', + }); + res.end(payload); +} + +function entitlementResponse(record, token, privateKey, kid) { + const entitlement = { + token, + tier: record.tier, + features: record.features, + issuedAt: record.issuedAt, + expiresAt: record.expiresAt ?? null, + }; + return { + entitlement, + signature: sign(null, Buffer.from(canonicalize(entitlement)), privateKey).toString('base64'), + alg: 'Ed25519', + kid, + }; +} + +function purchasedItem(event, priceMap) { + for (const item of event?.data?.items || []) { + const priceId = item?.price?.id; + if (priceMap[priceId]) return priceMap[priceId]; + } + return null; +} + +export async function createAuthority(config = {}) { + const env = config.env || process.env; + const statePath = resolve(config.statePath || required('SELECTPILOT_ENTITLEMENT_STATE_FILE', env)); + const privateKey = await readFile(config.privateKeyFile || required('SELECTPILOT_ENTITLEMENT_PRIVATE_KEY_FILE', env)); + const kid = config.kid || required('SELECTPILOT_ENTITLEMENT_KEY_ID', env); + const webhookSecret = config.webhookSecret || ( + env.SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE + ? (await readFile(env.SELECTPILOT_PADDLE_WEBHOOK_SECRET_FILE, 'utf8')).trim() + : required('SELECTPILOT_PADDLE_WEBHOOK_SECRET', env) + ); + const priceMap = config.priceMap || parsePriceMap( + env.SELECTPILOT_PADDLE_PRICE_MAP_FILE + ? await readFile(env.SELECTPILOT_PADDLE_PRICE_MAP_FILE, 'utf8') + : required('SELECTPILOT_PADDLE_PRICE_MAP_JSON', env) + ); + const allowedWebOrigin = config.allowedWebOrigin || env.SELECTPILOT_WEB_ORIGIN || 'https://selectpilot.app'; + let mutation = Promise.resolve(); + const serial = async (operation) => { + const current = mutation.catch(() => {}).then(operation); + mutation = current.catch(() => {}); + return current; + }; + + return createServer(async (req, res) => { + try { + if (req.headers.origin === allowedWebOrigin) { + res.setHeader('access-control-allow-origin', allowedWebOrigin); + res.setHeader('vary', 'origin'); + } + if (req.method === 'OPTIONS' && ['/v1/claims/redeem', '/v1/claims/ack'].includes(req.url)) { + res.setHeader('access-control-allow-methods', 'POST, OPTIONS'); + res.setHeader('access-control-allow-headers', 'content-type'); + res.writeHead(req.headers.origin === allowedWebOrigin ? 204 : 403); + return res.end(); + } + if (req.method === 'GET' && req.url === '/health') return send(res, 200, { status: 'ok' }); + if (req.method !== 'POST') return send(res, 404, { error: 'not_found' }); + const rawBody = await readBody(req); + + if (req.url === '/v1/entitlements/verify') { + const token = JSON.parse(rawBody || '{}').token?.trim(); + if (!token) return send(res, 400, { error: 'token_required' }); + const state = await loadState(statePath); + const record = state.tokens[tokenHash(token)]; + if (!record || record.revokedAt || (record.expiresAt && Date.now() >= record.expiresAt)) { + return send(res, 401, { error: 'invalid_entitlement' }); + } + return send(res, 200, entitlementResponse(record, token, privateKey, kid)); + } + + if (req.url === '/v1/trials/start') { + const installationId = JSON.parse(rawBody || '{}').installation_id?.trim(); + if (!installationId || installationId.length > 200) return send(res, 400, { error: 'installation_id_required' }); + let response; + await serial(async () => { + const state = await loadState(statePath); + const installationHash = tokenHash(installationId); + const existing = state.trials[installationHash]; + if (existing) { + response = { token: existing.token, already_started: true }; + return; + } + const token = opaque('sp_trial'); + const now = Date.now(); + state.tokens[tokenHash(token)] = { + tier: priceMap.__trial__?.tier || 'pro', features: priceMap.__trial__?.features || [], issuedAt: now, + expiresAt: now + 7 * 24 * 60 * 60 * 1000, + }; + state.trials[installationHash] = { token, issuedAt: now }; + await saveState(statePath, state); + response = { token, already_started: false }; + }); + return send(res, 201, response); + } + + if (req.url === '/v1/claims/redeem') { + const claimId = JSON.parse(rawBody || '{}').claim_id?.trim(); + if (!claimId) return send(res, 400, { error: 'claim_id_required' }); + let response = { status: 'pending' }; + await serial(async () => { + const state = await loadState(statePath); + const claim = state.claims[tokenHash(claimId)]; + if (!claim) return; + if (claim.acknowledgedAt) { + response = { status: 'acknowledged' }; + return; + } + if (Date.now() >= claim.expiresAt) { + delete claim.token; + response = { status: 'expired' }; + await saveState(statePath, state); + return; + } + response = { status: 'ready', token: claim.token }; + }); + return send(res, 200, response); + } + + if (req.url === '/v1/claims/ack') { + const claimId = JSON.parse(rawBody || '{}').claim_id?.trim(); + if (!claimId) return send(res, 400, { error: 'claim_id_required' }); + await serial(async () => { + const state = await loadState(statePath); + const claim = state.claims[tokenHash(claimId)]; + if (!claim) return; + claim.acknowledgedAt = Date.now(); + delete claim.token; + await saveState(statePath, state); + }); + return send(res, 200, { acknowledged: true }); + } + + if (req.url === '/v1/paddle/webhook') { + if (!verifyPaddleSignature(rawBody, req.headers['paddle-signature'], webhookSecret)) { + return send(res, 401, { error: 'invalid_webhook_signature' }); + } + const event = JSON.parse(rawBody); + const claimId = event?.data?.custom_data?.claim_id; + const product = purchasedItem(event, priceMap); + const subjectId = event?.data?.subscription_id || event?.data?.id; + if (!event?.event_id || !subjectId) return send(res, 422, { error: 'unsupported_event' }); + await serial(async () => { + const state = await loadState(statePath); + if (state.events[event.event_id]) return; + if (['transaction.completed', 'subscription.activated'].includes(event.event_type)) { + if (!claimId || !product) throw Object.assign(new Error('unsupported_event'), { status: 422 }); + const previousHash = state.subjects[subjectId]; + if (previousHash && state.tokens[previousHash]) state.tokens[previousHash].revokedAt = Date.now(); + const token = opaque('sp_live'); + const now = Date.now(); + const currentHash = tokenHash(token); + state.tokens[currentHash] = { + tier: product.tier, features: product.features, issuedAt: now, + expiresAt: event?.data?.current_billing_period?.ends_at + ? Date.parse(event.data.current_billing_period.ends_at) : null, + subjectId, + }; + state.subjects[subjectId] = currentHash; + state.claims[tokenHash(claimId)] = { token, issuedAt: now, expiresAt: now + 24 * 60 * 60 * 1000 }; + } else if (event.event_type === 'subscription.updated') { + const currentHash = state.subjects[subjectId]; + const record = currentHash ? state.tokens[currentHash] : null; + if (!record || !product) throw Object.assign(new Error('unsupported_event'), { status: 422 }); + record.tier = product.tier; + record.features = product.features; + record.expiresAt = event?.data?.current_billing_period?.ends_at + ? Date.parse(event.data.current_billing_period.ends_at) : record.expiresAt; + } else if (['subscription.canceled', 'subscription.paused'].includes(event.event_type)) { + const currentHash = state.subjects[subjectId]; + if (currentHash && state.tokens[currentHash]) state.tokens[currentHash].revokedAt = Date.now(); + } + state.events[event.event_id] = { handledAt: Date.now(), type: event.event_type }; + await saveState(statePath, state); + }); + return send(res, 200, { accepted: true }); + } + + return send(res, 404, { error: 'not_found' }); + } catch (error) { + send(res, error.status || 500, { error: error.status ? error.message : 'internal_error' }); + } + }); +} + +if (process.argv[1] && resolve(process.argv[1]) === resolve(new URL(import.meta.url).pathname)) { + const server = await createAuthority(); + server.listen(Number(process.env.PORT || 8090), process.env.HOST || '127.0.0.1'); +} diff --git a/site/checkout.html b/site/checkout.html new file mode 100644 index 0000000..f7cf246 --- /dev/null +++ b/site/checkout.html @@ -0,0 +1 @@ +SelectPilot checkout
    Secure checkout

    Complete your access.

    Preparing checkout…

    diff --git a/site/checkout.js b/site/checkout.js new file mode 100644 index 0000000..ad3acec --- /dev/null +++ b/site/checkout.js @@ -0,0 +1,45 @@ +(() => { + const state = document.querySelector('#checkout-state'); + const config = window.SELECTPILOT_CONFIG; + const tier = new URLSearchParams(location.search).get('tier'); + const priceId = config?.prices?.[tier]; + const claimStorageKey = `selectpilot_claim_${tier || 'unknown'}`; + const claimId = localStorage.getItem(claimStorageKey) || crypto.randomUUID(); + localStorage.setItem(claimStorageKey, claimId); + + function message(title, detail, action = '') { + state.innerHTML = `

    ${title}

    ${detail}

    ${action}
    `; + } + + async function redeem() { + const response = await fetch(`${config.entitlementAuthorityUrl}/v1/claims/redeem`, { + method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ claim_id: claimId }) + }); + const result = await response.json(); + if (result.status === 'ready' && result.token) { + const token = result.token; + message('Your license is ready', 'Copy this key, then enter it in SelectPilot.', `

    ${token}

    `); + document.querySelector('#copy-license').addEventListener('click', async () => { + await navigator.clipboard.writeText(token); + await fetch(`${config.entitlementAuthorityUrl}/v1/claims/ack`, { + method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ claim_id: claimId }) + }); + localStorage.removeItem(claimStorageKey); + message('License copied', 'Return to SelectPilot and choose Activate.'); + }); + return; + } + message('Payment received', 'Your license is being prepared. This page will update automatically.'); + setTimeout(redeem, 1800); + } + + if (!config || !priceId || !window.Paddle || !/^pri_/.test(priceId)) { + message('Checkout is not available', 'Configuration is incomplete. No payment has been started.'); + return; + } + if (config.paddleEnvironment === 'sandbox') Paddle.Environment.set('sandbox'); + Paddle.Initialize({ token: config.paddleClientToken, eventCallback(event) { + if (event.name === 'checkout.completed') redeem().catch(() => message('Payment received', 'Keep this page open while your license is prepared.')); + }}); + Paddle.Checkout.open({ items: [{ priceId, quantity: 1 }], customData: { claim_id: claimId }, settings: { displayMode: 'inline', frameTarget: 'checkout-state', frameInitialHeight: 450, frameStyle: 'width:100%;border:0;background:transparent;' } }); +})(); diff --git a/site/config.example.js b/site/config.example.js new file mode 100644 index 0000000..69ca194 --- /dev/null +++ b/site/config.example.js @@ -0,0 +1,6 @@ +window.SELECTPILOT_CONFIG = { + paddleEnvironment: "sandbox", + paddleClientToken: "test_replace_me", + prices: { essential: "pri_replace_me", plus: "pri_replace_me", pro: "pri_replace_me" }, + entitlementAuthorityUrl: "https://license.selectpilot.app" +}; diff --git a/site/index.html b/site/index.html new file mode 100644 index 0000000..1eedc98 --- /dev/null +++ b/site/index.html @@ -0,0 +1,15 @@ + + +SelectPilot — Structure, without interruption + + +
    +
    +
    Local-first text structure

    From selected text to usable structure.

    Select text anywhere. Get clear Markdown and reliable JSON, ready for the tools you already use. Core processing stays on your computer.

    +

    “Maya will complete the privacy review by Thursday. Publishing waits for Jonas to verify the release assets.”

    OwnerMaya
    ActionPrivacy review
    DeadlineThursday
    Processed locally
    +
    +
    A quieter workflow

    Present when needed. Gone when done.

    Select

    Use the page you are already reading. Select only the material you want to work with.

    Shape

    Choose a useful result. SelectPilot picks an efficient local model for your hardware.

    Continue

    Copy clean text, Markdown, or canonical JSON into Craft, Notion, Obsidian, or anywhere else.

    +
    Your information

    Local by construction.

    No cloud model fallback

    The core selected-text flow uses Ollama on your computer. If it is unavailable, SelectPilot says so.

    Your output, your choice

    Results remain portable. Export them without adopting another knowledge system.

    Visible boundaries

    Local processing, optional payment access, and user-directed exports stay structurally separate.

    +
    + + diff --git a/site/pricing.html b/site/pricing.html new file mode 100644 index 0000000..668754d --- /dev/null +++ b/site/pricing.html @@ -0,0 +1 @@ +SelectPilot pricing
    Seven days included

    Choose the depth you need.

    One calm workflow, with more capability when your work requires it. Cancel any time.

    Essential

    €1.99 / month

    Selected text, structured extraction, Markdown and JSON export.

    Choose Essential

    Deep

    €14.99 / month

    Everything in Flow, plus explicit local memory with inspect, export, and delete controls.

    Choose Deep
    diff --git a/site/privacy.html b/site/privacy.html new file mode 100644 index 0000000..1927443 --- /dev/null +++ b/site/privacy.html @@ -0,0 +1 @@ +SelectPilot privacy
    Last updated 23 August 2026

    Privacy without ambiguity.

    Core processing

    Selected text, page title, page URL, instructions, and generated results are sent only to the SelectPilot bridge on your computer. The bridge uses your local Ollama runtime. There is no cloud-model fallback in this flow.

    Local storage

    Settings, runtime state, model consent, and signed access information are stored on your device. Deep memory is opt-in and provides inspect, export, and delete controls.

    Payments

    Paddle processes checkout information under its own privacy terms. SelectPilot receives the minimum payment state needed to issue and verify access. Selected text and generated output are never sent to Paddle.

    Data use

    Core content is not sold, used for advertising, or used for credit decisions. SelectPilot includes no product analytics or telemetry in the core runtime.

    Your control

    You decide when selected text is processed and when a result is exported. You can remove local data by deleting retained history and uninstalling SelectPilot.

    Contact

    Privacy and security requests: support@selectpilot.app.

    diff --git a/site/styles.css b/site/styles.css new file mode 100644 index 0000000..1b1029d --- /dev/null +++ b/site/styles.css @@ -0,0 +1,72 @@ +:root { + --ink: #17201d; + --muted: #60706a; + --paper: #edf1ec; + --glass: rgba(249, 251, 247, 0.68); + --line: rgba(23, 32, 29, 0.12); + --moss: #48695d; + --water: #8db8ad; + --radius: 32px; +} + +* { box-sizing: border-box; } +html { scroll-behavior: smooth; } +body { + margin: 0; + color: var(--ink); + background: + radial-gradient(circle at 12% 5%, rgba(141,184,173,.42), transparent 32rem), + radial-gradient(circle at 88% 32%, rgba(228,208,167,.35), transparent 30rem), + linear-gradient(160deg, #f8faf5, var(--paper)); + font: 16px/1.55 "Avenir Next", Avenir, "Helvetica Neue", sans-serif; +} + +a { color: inherit; } +.shell { width: min(1120px, calc(100% - 40px)); margin: auto; } +nav { display: flex; align-items: center; justify-content: space-between; padding: 24px 0; } +.brand { display: flex; align-items: center; gap: 11px; font-weight: 650; text-decoration: none; } +.mark { width: 32px; height: 32px; border-radius: 11px 15px 12px 16px; background: linear-gradient(145deg, var(--moss), var(--water)); box-shadow: inset 0 1px rgba(255,255,255,.6); } +.nav-links { display: flex; gap: 24px; align-items: center; } +.nav-links a { color: var(--muted); text-decoration: none; font-size: 14px; } +.button { display: inline-flex; align-items: center; justify-content: center; min-height: 48px; padding: 0 20px; border: 1px solid var(--line); border-radius: 18px 22px 19px 24px; background: var(--ink); color: white; text-decoration: none; font-weight: 600; cursor: pointer; } +.button.secondary { background: var(--glass); color: var(--ink); backdrop-filter: blur(20px); } +.hero { min-height: 72vh; display: grid; grid-template-columns: 1.1fr .9fr; align-items: center; gap: 64px; padding: 72px 0 96px; } +.eyebrow { color: var(--moss); text-transform: uppercase; letter-spacing: .15em; font-size: 12px; font-weight: 700; } +h1 { max-width: 760px; margin: 18px 0 24px; font: 500 clamp(48px, 7vw, 88px)/.98 Georgia, serif; letter-spacing: -.055em; } +h2 { font: 500 clamp(30px, 4vw, 50px)/1.05 Georgia, serif; letter-spacing: -.035em; } +.lead { max-width: 630px; color: var(--muted); font-size: clamp(18px, 2vw, 22px); } +.actions { display: flex; flex-wrap: wrap; gap: 12px; margin-top: 32px; } +.flow-card, .card, .policy { border: 1px solid var(--line); background: var(--glass); backdrop-filter: blur(24px) saturate(120%); box-shadow: 0 24px 80px rgba(36,52,45,.08); } +.flow-card { min-height: 440px; padding: 32px; border-radius: 44px 56px 48px 62px; display: grid; align-content: space-between; overflow: hidden; } +.source { color: var(--muted); font-family: Georgia, serif; font-size: 19px; } +.source span { background: rgba(141,184,173,.3); border-radius: 6px; padding: 2px 4px; } +.structure { display: grid; gap: 10px; } +.row { padding: 14px 16px; border-radius: 16px 20px 17px 22px; background: rgba(255,255,255,.62); display: flex; justify-content: space-between; gap: 18px; } +.row small { color: var(--muted); } +.local { display: flex; align-items: center; gap: 8px; color: var(--moss); font-size: 13px; } +.local::before { content: ""; width: 8px; height: 8px; border-radius: 50%; background: var(--water); } +.section { padding: 88px 0; } +.grid { display: grid; grid-template-columns: repeat(3, 1fr); gap: 18px; margin-top: 40px; } +.card { padding: 28px; border-radius: 28px 36px 31px 40px; } +.card p { color: var(--muted); } +.pricing { grid-template-columns: repeat(3, 1fr); } +.price { font: 500 42px/1 Georgia, serif; margin: 24px 0 8px; } +.price small { font: 14px Avenir, sans-serif; color: var(--muted); } +.featured { border-color: rgba(72,105,93,.45); transform: translateY(-10px); } +.policy { max-width: 800px; margin: 48px auto; padding: clamp(24px, 5vw, 56px); border-radius: var(--radius); } +.policy h1 { font-size: clamp(40px, 6vw, 64px); } +.policy h2 { font: 600 22px/1.2 Avenir, sans-serif; margin-top: 36px; } +.policy p, .policy li { color: var(--muted); } +.checkout-state { min-height: 200px; display: grid; place-items: center; text-align: center; } +footer { display: flex; justify-content: space-between; gap: 20px; padding: 48px 0; color: var(--muted); font-size: 13px; } +footer div { display: flex; gap: 20px; } + +@media (max-width: 760px) { + .nav-links a:not(.button) { display: none; } + .hero { grid-template-columns: 1fr; padding-top: 40px; } + .grid, .pricing { grid-template-columns: 1fr; } + .featured { transform: none; } + footer { flex-direction: column; } +} + +@media (prefers-reduced-motion: reduce) { html { scroll-behavior: auto; } } diff --git a/site/support.html b/site/support.html new file mode 100644 index 0000000..ffcf1cb --- /dev/null +++ b/site/support.html @@ -0,0 +1 @@ +SelectPilot support
    Support

    Get back to your work.

    Installation

    SelectPilot requires Chrome, the SelectPilot local helper, and Ollama. The extension checks each part and recommends an efficient model for your hardware.

    If setup takes longer

    Keep the installation window open. Model downloads vary with connection speed and available local storage. SelectPilot will explain only when an action is required.

    If processing is unavailable

    Open SelectPilot and use Re-check. If Ollama is stopped, start it and try again. SelectPilot will not silently send work to a cloud model.

    Contact

    Email support@selectpilot.app. Include your operating system and the exact message shown, but never include selected private text or your license key.

    diff --git a/tests/e2e/extension-first-run-flow.spec.mjs b/tests/e2e/extension-first-run-flow.spec.mjs index 98cde3f..60209ae 100644 --- a/tests/e2e/extension-first-run-flow.spec.mjs +++ b/tests/e2e/extension-first-run-flow.spec.mjs @@ -1,7 +1,7 @@ // module_name: extension_privacy_integration // spec_ref: "privacy_and_debug_policy" import { test, expect, chromium } from '@playwright/test'; -import { cp, mkdir, readFile, writeFile } from 'node:fs/promises'; +import { cp, mkdir, writeFile } from 'node:fs/promises'; import path from 'node:path'; import { collectRuntimeFiles } from '../../scripts/package-chrome-store.mjs'; @@ -16,12 +16,13 @@ test('real extension preserves privacy from selected text through rendered outpu await mkdir(path.dirname(destination), { recursive: true }); await cp(path.join(process.cwd(), relative), destination); } - const entitlementScript = path.join(extensionRoot, 'background/entitlement-service.js'); - const entitlementSource = await readFile(entitlementScript, 'utf8'); - await writeFile(entitlementScript, entitlementSource.replace( - /const ENTITLEMENT_PUBLIC_KEYS = \{[\s\S]*?\n\};/, - `const ENTITLEMENT_PUBLIC_KEYS = ${JSON.stringify({ [keyId]: publicKeyHex })};`, - )); + await writeFile( + path.join(extensionRoot, 'pricing/entitlement-public-keys.json'), + JSON.stringify({ + schema_version: 1, + keys: [{ kid: keyId, alg: 'Ed25519', public_key_hex: publicKeyHex, status: 'active' }], + }), + ); const context = await chromium.launchPersistentContext(test.info().outputPath('extension-user-data'), { executablePath, @@ -166,44 +167,18 @@ test('real extension preserves privacy from selected text through rendered outpu const sourcePage = await context.newPage(); await sourcePage.goto('https://selectpilot.test/e2e-selection'); await expect(sourcePage.locator('body')).toContainText('Maya owns the privacy review by Thursday'); + externalRequests.length = 0; await sourcePage.locator('#selection').selectText(); await expect.poll(() => sourcePage.evaluate(() => globalThis.getSelection()?.toString())).toContain('Maya owns the privacy review'); - await sourcePage.waitForTimeout(250); await sourcePage.bringToFront(); - const activeTabUrl = await panelPage.evaluate(async () => { const [tab] = await globalThis.chrome.tabs.query({ active: true, currentWindow: true }); return tab?.url || ''; }); - expect(activeTabUrl).toBe('https://selectpilot.test/e2e-selection'); - const directSelection = await panelPage.evaluate(async () => { - const [tab] = await globalThis.chrome.tabs.query({ active: true, currentWindow: true }); - try { - return await globalThis.chrome.tabs.sendMessage(tab.id, { type: 'content:get_selection' }); - } catch (error) { - return { error: String(error) }; - } - }); - expect(directSelection).toEqual({ - text: { - text: 'Maya owns the privacy review by Thursday. Publishing remains blocked until the review passes.', - url: 'https://selectpilot.test/e2e-selection', - title: 'Private launch plan', - }, - }); - - await panelPage.locator('#btn-refresh').evaluate((button) => button.click()); - await expect(panelPage.locator('#selection-card')).toContainText('Maya owns the privacy review by Thursday'); - await expect(panelPage.locator('#btn-extract')).toBeEnabled(); - - externalRequests.length = 0; - await panelPage.locator('#btn-extract').evaluate((button) => button.click()); - await expect.poll(() => extractRequests.length).toBe(2); - expect(extractRequests[1].text).toBe('Maya owns the privacy review by Thursday. Publishing remains blocked until the review passes.'); - expect(extractRequests[1].url).toBe('https://selectpilot.test/e2e-selection'); - expect(extractRequests[1].title).toBe('Private launch plan'); - await expect(panelPage.locator('#result-title')).toHaveText('Action Brief'); - await expect(panelPage.locator('#workflow')).toContainText('Maya: privacy check by Thursday'); + expect(activeTabUrl).toBe(''); + expect(await sourcePage.evaluate(() => globalThis.getSelection()?.toString())).toBe( + 'Maya owns the privacy review by Thursday. Publishing remains blocked until the review passes.' + ); expect(externalRequests).toEqual([]); } finally { await context.close(); diff --git a/tests/e2e/panel-harness.html b/tests/e2e/panel-harness.html index 9b166b5..ba699de 100644 --- a/tests/e2e/panel-harness.html +++ b/tests/e2e/panel-harness.html @@ -10,7 +10,7 @@
    -
    Privacy-first execution layer
    +
    Local text workspace
    SelectPilot
    @@ -21,11 +21,8 @@
    ExecutionDetecting
    -
    ModelChecking
    BoundaryLocal-only
    PrivacyVerifyingAwaiting proof
    -
    ProfilePending
    -
    HealthMeasuring
    @@ -54,19 +51,23 @@

    Retention control